feat(policies): add Nvidia Gr00t N1.5 model (#2292)

* feat(policies): add Nvidia Gr00t N1.5 model

Co-authored-by: lbenhorin <lbenhorin@nvidia.com>
Co-authored-by: Aravindh <aravindhs@nvidia.com>
Co-authored-by: nv-sachdevkartik <ksachdev@nvidia.com>
Co-authored-by: youliangt <youliangt@nvidia.com>
Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>
Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>
Co-authored-by: Jade Choghari <chogharijade@gmail.com>

* fix(docs): add groot to index

Co-authored-by: sachdevkartik <sachdev.kartik25@gmail.com>

---------

Co-authored-by: lbenhorin <lbenhorin@nvidia.com>
Co-authored-by: Aravindh <aravindhs@nvidia.com>
Co-authored-by: nv-sachdevkartik <ksachdev@nvidia.com>
Co-authored-by: youliangt <youliangt@nvidia.com>
Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>
Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>
Co-authored-by: Jade Choghari <chogharijade@gmail.com>
Co-authored-by: sachdevkartik <sachdev.kartik25@gmail.com>

src/lerobot/policies/__init__.py

@@ -14,6 +14,7 @@
 
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
+from .groot.configuration_groot import GrootConfig as GrootConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .pi05.configuration_pi05 import PI05Config as PI05Config
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
@@ -29,4 +30,5 @@ __all__ = [
     "SmolVLAConfig",
     "TDMPCConfig",
     "VQBeTConfig",
+    "GrootConfig",
 ]

src/lerobot/policies/factory.py

@@ -30,6 +30,7 @@ from lerobot.envs.configs import EnvConfig
 from lerobot.envs.utils import env_to_policy_features
 from lerobot.policies.act.configuration_act import ACTConfig
 from lerobot.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.policies.groot.configuration_groot import GrootConfig
 from lerobot.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.policies.pi05.configuration_pi05 import PI05Config
 from lerobot.policies.pretrained import PreTrainedPolicy
@@ -101,6 +102,10 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
         from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
 
         return SmolVLAPolicy
+    elif name == "groot":
+        from lerobot.policies.groot.modeling_groot import GrootPolicy
+
+        return GrootPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -142,6 +147,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return SmolVLAConfig(**kwargs)
     elif policy_type == "reward_classifier":
         return RewardClassifierConfig(**kwargs)
+    elif policy_type == "groot":
+        return GrootConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
@@ -199,6 +206,27 @@ def make_pre_post_processors(
             policy configuration type.
     """
     if pretrained_path:
+        # TODO(Steven): Temporary patch, implement correctly the processors for Gr00t
+        if isinstance(policy_cfg, GrootConfig):
+            # GROOT handles normalization in groot_pack_inputs_v3 step
+            # Need to override both stats AND normalize_min_max since saved config might be empty
+            preprocessor_overrides = {}
+            postprocessor_overrides = {}
+            preprocessor_overrides["groot_pack_inputs_v3"] = {
+                "stats": kwargs.get("dataset_stats"),
+                "normalize_min_max": True,
+            }
+
+            # Also ensure postprocessing slices to env action dim and unnormalizes with dataset stats
+            env_action_dim = policy_cfg.output_features["action"].shape[0]
+            postprocessor_overrides["groot_action_unpack_unnormalize_v1"] = {
+                "stats": kwargs.get("dataset_stats"),
+                "normalize_min_max": True,
+                "env_action_dim": env_action_dim,
+            }
+            kwargs["preprocessor_overrides"] = preprocessor_overrides
+            kwargs["postprocessor_overrides"] = postprocessor_overrides
+
         return (
             PolicyProcessorPipeline.from_pretrained(
                 pretrained_model_name_or_path=pretrained_path,
@@ -293,6 +321,14 @@ def make_pre_post_processors(
             dataset_stats=kwargs.get("dataset_stats"),
         )
 
+    elif isinstance(policy_cfg, GrootConfig):
+        from lerobot.policies.groot.processor_groot import make_groot_pre_post_processors
+
+        processors = make_groot_pre_post_processors(
+            config=policy_cfg,
+            dataset_stats=kwargs.get("dataset_stats"),
+        )
+
     else:
         raise NotImplementedError(f"Processor for policy type '{policy_cfg.type}' is not implemented.")
 

src/lerobot/policies/groot/README.md

@@ -0,0 +1 @@
+../../../../docs/source/policy_groot_README.md

src/lerobot/policies/groot/__init__.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Nvidia and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_groot import GrootConfig
+from .modeling_groot import GrootPolicy
+from .processor_groot import make_groot_pre_post_processors
+
+__all__ = ["GrootConfig", "GrootPolicy", "make_groot_pre_post_processors"]

src/lerobot/policies/groot/action_head/__init__.py

@@ -0,0 +1,14 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

src/lerobot/policies/groot/action_head/action_encoder.py

@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+
+
+def swish(x):
+    return x * torch.sigmoid(x)
+
+
+class SinusoidalPositionalEncoding(nn.Module):
+    """
+    Produces a sinusoidal encoding of shape (B, T, w)
+    given timesteps of shape (B, T).
+    """
+
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+
+    def forward(self, timesteps):
+        # timesteps: shape (B, T)
+        # We'll compute sin/cos frequencies across dim T
+        timesteps = timesteps.float()  # ensure float
+
+        b, t = timesteps.shape
+        device = timesteps.device
+
+        half_dim = self.embedding_dim // 2
+        # typical log space frequencies for sinusoidal encoding
+        exponent = -torch.arange(half_dim, dtype=torch.float, device=device) * (
+            torch.log(torch.tensor(10000.0)) / half_dim
+        )
+        # Expand timesteps to (B, T, 1) then multiply
+        freqs = timesteps.unsqueeze(-1) * exponent.exp()  # (B, T, half_dim)
+
+        sin = torch.sin(freqs)
+        cos = torch.cos(freqs)
+        enc = torch.cat([sin, cos], dim=-1)  # (B, T, w)
+
+        return enc

src/lerobot/policies/groot/action_head/cross_attention_dit.py

@@ -0,0 +1,370 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from diffusers import ConfigMixin, ModelMixin
+from diffusers.configuration_utils import register_to_config
+from diffusers.models.attention import Attention, FeedForward
+from diffusers.models.embeddings import (
+    SinusoidalPositionalEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from torch import nn
+
+
+class TimestepEncoder(nn.Module):
+    def __init__(self, embedding_dim, compute_dtype=torch.float32):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(self, timesteps):
+        dtype = next(self.parameters()).dtype
+        timesteps_proj = self.time_proj(timesteps).to(dtype)
+        timesteps_emb = self.timestep_embedder(timesteps_proj)  # (N, D)
+        return timesteps_emb
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        chunk_dim: int = 0,
+    ):
+        super().__init__()
+        self.chunk_dim = chunk_dim
+        output_dim = embedding_dim * 2
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, output_dim)
+        self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        temb: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        temb = self.linear(self.silu(temb))
+        scale, shift = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale[:, None]) + shift[:, None]
+        return x
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: int | None = None,
+        activation_fn: str = "geglu",
+        attention_bias: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: str | None = None,
+        num_positional_embeddings: int | None = None,
+        ff_inner_dim: int | None = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.norm_type = norm_type
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embeddings` type is defined, `num_positional_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if norm_type == "ada_norm":
+            self.norm1 = AdaLayerNorm(dim)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+        if final_dropout:
+            self.final_dropout = nn.Dropout(dropout)
+        else:
+            self.final_dropout = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        encoder_hidden_states: torch.Tensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        temb: torch.LongTensor | None = None,
+    ) -> torch.Tensor:
+        # 0. Self-Attention
+        if self.norm_type == "ada_norm":
+            norm_hidden_states = self.norm1(hidden_states, temb)
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            # encoder_attention_mask=encoder_attention_mask,
+        )
+        if self.final_dropout:
+            attn_output = self.final_dropout(attn_output)
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+        ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+        return hidden_states
+
+
+class DiT(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 8,
+        attention_head_dim: int = 64,
+        output_dim: int = 26,
+        num_layers: int = 12,
+        dropout: float = 0.1,
+        attention_bias: bool = True,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: int | None = 1000,
+        upcast_attention: bool = False,
+        norm_type: str = "ada_norm",
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        max_num_positional_embeddings: int = 512,
+        compute_dtype=torch.float32,
+        final_dropout: bool = True,
+        positional_embeddings: str | None = "sinusoidal",
+        interleave_self_attention=False,
+        cross_attention_dim: int | None = None,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.gradient_checkpointing = False
+
+        # Timestep encoder
+        self.timestep_encoder = TimestepEncoder(
+            embedding_dim=self.inner_dim, compute_dtype=self.config.compute_dtype
+        )
+
+        all_blocks = []
+        for idx in range(self.config.num_layers):
+            use_self_attn = idx % 2 == 1 and interleave_self_attention
+            curr_cross_attention_dim = cross_attention_dim if not use_self_attn else None
+
+            all_blocks += [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    activation_fn=self.config.activation_fn,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    positional_embeddings=positional_embeddings,
+                    num_positional_embeddings=self.config.max_num_positional_embeddings,
+                    final_dropout=final_dropout,
+                    cross_attention_dim=curr_cross_attention_dim,
+                )
+            ]
+        self.transformer_blocks = nn.ModuleList(all_blocks)
+
+        # Output blocks
+        self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim)
+        self.proj_out_2 = nn.Linear(self.inner_dim, self.config.output_dim)
+        print(
+            "Total number of DiT parameters: ",
+            sum(p.numel() for p in self.parameters() if p.requires_grad),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,  # Shape: (B, T, D)
+        encoder_hidden_states: torch.Tensor,  # Shape: (B, S, D)
+        timestep: torch.LongTensor | None = None,
+        encoder_attention_mask: torch.Tensor | None = None,
+        return_all_hidden_states: bool = False,
+    ):
+        # Encode timesteps
+        temb = self.timestep_encoder(timestep)
+
+        # Process through transformer blocks - single pass through the blocks
+        hidden_states = hidden_states.contiguous()
+        encoder_hidden_states = encoder_hidden_states.contiguous()
+
+        all_hidden_states = [hidden_states]
+
+        # Process through transformer blocks
+        for idx, block in enumerate(self.transformer_blocks):
+            if idx % 2 == 1 and self.config.interleave_self_attention:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=None,
+                    encoder_hidden_states=None,
+                    encoder_attention_mask=None,
+                    temb=temb,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=None,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=None,
+                    temb=temb,
+                )
+            all_hidden_states.append(hidden_states)
+
+        # Output processing
+        conditioning = temb
+        shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+        if return_all_hidden_states:
+            return self.proj_out_2(hidden_states), all_hidden_states
+        else:
+            return self.proj_out_2(hidden_states)
+
+
+class SelfAttentionTransformer(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 8,
+        attention_head_dim: int = 64,
+        output_dim: int = 26,
+        num_layers: int = 12,
+        dropout: float = 0.1,
+        attention_bias: bool = True,
+        activation_fn: str = "gelu-approximate",
+        num_embeds_ada_norm: int | None = 1000,
+        upcast_attention: bool = False,
+        max_num_positional_embeddings: int = 512,
+        compute_dtype=torch.float32,
+        final_dropout: bool = True,
+        positional_embeddings: str | None = "sinusoidal",
+        interleave_self_attention=False,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.gradient_checkpointing = False
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    activation_fn=self.config.activation_fn,
+                    attention_bias=self.config.attention_bias,
+                    upcast_attention=self.config.upcast_attention,
+                    positional_embeddings=positional_embeddings,
+                    num_positional_embeddings=self.config.max_num_positional_embeddings,
+                    final_dropout=final_dropout,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+        print(
+            "Total number of SelfAttentionTransformer parameters: ",
+            sum(p.numel() for p in self.parameters() if p.requires_grad),
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,  # Shape: (B, T, D)
+        return_all_hidden_states: bool = False,
+    ):
+        # Process through transformer blocks - single pass through the blocks
+        hidden_states = hidden_states.contiguous()
+        all_hidden_states = [hidden_states]
+
+        # Process through transformer blocks
+        for _idx, block in enumerate(self.transformer_blocks):
+            hidden_states = block(hidden_states)
+            all_hidden_states.append(hidden_states)
+
+        if return_all_hidden_states:
+            return hidden_states, all_hidden_states
+        else:
+            return hidden_states

src/lerobot/policies/groot/action_head/flow_matching_action_head.py

@@ -0,0 +1,406 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import nn
+from torch.distributions import Beta
+
+from lerobot.utils.import_utils import _transformers_available
+
+# Conditional import for type checking and lazy loading
+if TYPE_CHECKING or _transformers_available:
+    from transformers import PretrainedConfig
+    from transformers.feature_extraction_utils import BatchFeature
+else:
+    PretrainedConfig = object
+    BatchFeature = None
+
+from lerobot.policies.groot.action_head.action_encoder import (
+    SinusoidalPositionalEncoding,
+    swish,
+)
+
+from .cross_attention_dit import DiT, SelfAttentionTransformer
+
+
+class CategorySpecificLinear(nn.Module):
+    def __init__(self, num_categories, input_dim, hidden_dim):
+        super().__init__()
+        self.num_categories = num_categories
+        # For each category, we have separate weights and biases.
+        self.W = nn.Parameter(0.02 * torch.randn(num_categories, input_dim, hidden_dim))
+        self.b = nn.Parameter(torch.zeros(num_categories, hidden_dim))
+
+    def forward(self, x, cat_ids):
+        selected_w = self.W[cat_ids]
+        selected_b = self.b[cat_ids]
+        return torch.bmm(x, selected_w) + selected_b.unsqueeze(1)
+
+
+class CategorySpecificMLP(nn.Module):
+    def __init__(self, num_categories, input_dim, hidden_dim, output_dim):
+        super().__init__()
+        self.num_categories = num_categories
+        self.layer1 = CategorySpecificLinear(num_categories, input_dim, hidden_dim)
+        self.layer2 = CategorySpecificLinear(num_categories, hidden_dim, output_dim)
+
+    def forward(self, x, cat_ids):
+        hidden = F.relu(self.layer1(x, cat_ids))
+        return self.layer2(hidden, cat_ids)
+
+
+class MultiEmbodimentActionEncoder(nn.Module):
+    def __init__(self, action_dim, hidden_size, num_embodiments):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.num_embodiments = num_embodiments
+
+        # W1: R^{w x d}, W2: R^{w x 2w}, W3: R^{w x w}
+        self.W1 = CategorySpecificLinear(num_embodiments, action_dim, hidden_size)  # (d -> w)
+        self.W2 = CategorySpecificLinear(num_embodiments, 2 * hidden_size, hidden_size)  # (2w -> w)
+        self.W3 = CategorySpecificLinear(num_embodiments, hidden_size, hidden_size)  # (w -> w)
+        self.pos_encoding = SinusoidalPositionalEncoding(hidden_size)
+
+    def forward(self, actions, timesteps, cat_ids):
+        """
+        actions:   shape (B, T, action_dim)
+        timesteps: shape (B,)  -- a single scalar per batch item
+        cat_ids:   shape (B,)
+        returns:   shape (B, T, hidden_size)
+        """
+        b, t, _ = actions.shape
+
+        # 1) Expand each batch's single scalar time 'tau' across all T steps
+        #    so that shape => (B, T)
+        #    e.g. if timesteps is (B,), replicate across T
+        if timesteps.dim() == 1 and timesteps.shape[0] == b:
+            # shape (B,) => (B,T)
+            timesteps = timesteps.unsqueeze(1).expand(-1, t)
+        else:
+            raise ValueError("Expected `timesteps` to have shape (B,) so we can replicate across T.")
+
+        # 2) Standard action MLP step for shape => (B, T, w)
+        a_emb = self.W1(actions, cat_ids)
+
+        # 3) Get the sinusoidal encoding (B, T, w)
+        tau_emb = self.pos_encoding(timesteps).to(dtype=a_emb.dtype)
+
+        # 4) Concat along last dim => (B, T, 2w), then W2 => (B, T, w), swish
+        x = torch.cat([a_emb, tau_emb], dim=-1)
+        x = swish(self.W2(x, cat_ids))
+
+        # 5) Finally W3 => (B, T, w)
+        x = self.W3(x, cat_ids)
+        return x
+
+
+@dataclass
+class FlowmatchingActionHeadConfig(PretrainedConfig):
+    """NOTE: N1.5 uses XEmbFlowmatchingPolicyHeadConfig as action head"""
+
+    add_pos_embed: bool = field(default=True, metadata={"help": "Whether to add positional embedding"})
+    model_dtype: str = field(default="float32", metadata={"help": "Model data type."})
+    diffusion_model_cfg: dict = field(default=None, metadata={"help": "Diffusion model configuration."})
+    input_embedding_dim: int = field(default=1536, metadata={"help": "Input embedding channel dimension."})
+    backbone_embedding_dim: int = field(
+        default=1536, metadata={"help": "Backbone embedding channel dimension."}
+    )
+
+    hidden_size: int = field(default=1024, metadata={"help": "Input embedding dimension."})
+    max_seq_len: int = field(default=1024, metadata={"help": "Maximum Sequence Length"})
+    action_dim: int = field(default=None, metadata={"help": "Action dimension."})
+    action_horizon: int = field(default=None, metadata={"help": "Action horizon."})
+    noise_beta_alpha: float = field(default=1.5, metadata={"help": ""})
+    noise_beta_beta: float = field(default=1.0, metadata={"help": ""})
+    noise_s: float = field(default=0.999, metadata={"help": "Flow matching noise Beta distribution s."})
+    num_timestep_buckets: int = field(
+        default=1000, metadata={"help": "Number of timestep discretization buckets."}
+    )
+    num_inference_timesteps: int = field(
+        default=None,
+        metadata={"help": "Number of inference steps for noise diffusion."},
+    )
+    max_num_embodiments: int = field(default=32, metadata={"help": "Number of embodiments."})
+    tune_projector: bool = field(default=True, metadata={"help": "Whether to tune the projector."})
+    tune_diffusion_model: bool = field(
+        default=True, metadata={"help": "Whether to tune the diffusion model."}
+    )
+    load_pretrained_det_decode_layer_path: str = field(
+        default=None, metadata={"help": "Path to pretrained detection model."}
+    )
+    detection_coeff: float = field(default=1.0, metadata={"help": "Detection coefficient."})
+
+    freeze_decode_layer: bool = field(default=False)
+    expand_batch: int = field(default=None)
+    use_vlln: bool = field(default=True)
+
+    vl_self_attention_cfg: dict = field(default=None)
+    num_target_vision_tokens: int = field(default=32, metadata={"help": "Number of target vision tokens."})
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        for key, value in kwargs.items():
+            setattr(self, key, value)
+
+
+class FlowmatchingActionHead(nn.Module):
+    config_class = FlowmatchingActionHeadConfig
+    supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        config: FlowmatchingActionHeadConfig,
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.input_embedding_dim = config.input_embedding_dim
+
+        self.model = DiT(**config.diffusion_model_cfg)
+        self.action_dim = config.action_dim
+        self.action_horizon = config.action_horizon
+        self.num_inference_timesteps = config.num_inference_timesteps
+
+        self.state_encoder = CategorySpecificMLP(
+            num_categories=config.max_num_embodiments,
+            input_dim=config.max_state_dim,
+            hidden_dim=self.hidden_size,
+            output_dim=self.input_embedding_dim,
+        )
+        self.action_encoder = MultiEmbodimentActionEncoder(
+            action_dim=config.action_dim,
+            hidden_size=self.input_embedding_dim,
+            num_embodiments=config.max_num_embodiments,
+        )
+        self.action_decoder = CategorySpecificMLP(
+            num_categories=config.max_num_embodiments,
+            input_dim=self.hidden_size,
+            hidden_dim=self.hidden_size,
+            output_dim=self.action_dim,
+        )
+        self.future_tokens = nn.Embedding(config.num_target_vision_tokens, self.input_embedding_dim)
+        nn.init.normal_(self.future_tokens.weight, mean=0.0, std=0.02)
+
+        self.vlln = nn.LayerNorm(config.backbone_embedding_dim) if config.use_vlln else nn.Identity()
+        self.vl_self_attention = (
+            SelfAttentionTransformer(**config.vl_self_attention_cfg) if config.use_vlln else nn.Identity()
+        )
+
+        if config.add_pos_embed:
+            self.position_embedding = nn.Embedding(config.max_seq_len, self.input_embedding_dim)
+            nn.init.normal_(self.position_embedding.weight, mean=0.0, std=0.02)
+
+        self.beta_dist = Beta(config.noise_beta_alpha, config.noise_beta_beta)
+        self.num_timestep_buckets = config.num_timestep_buckets
+        self.config = config
+        self.set_trainable_parameters(config.tune_projector, config.tune_diffusion_model)
+
+    def set_trainable_parameters(self, tune_projector: bool, tune_diffusion_model: bool):
+        self.tune_projector = tune_projector
+        self.tune_diffusion_model = tune_diffusion_model
+        for p in self.parameters():
+            p.requires_grad = True
+        if not tune_projector:
+            self.state_encoder.requires_grad_(False)
+            self.action_encoder.requires_grad_(False)
+            self.action_decoder.requires_grad_(False)
+            if self.config.add_pos_embed:
+                self.position_embedding.requires_grad_(False)
+        if not tune_diffusion_model:
+            self.model.requires_grad_(False)
+        print(f"Tune action head projector: {self.tune_projector}")
+        print(f"Tune action head diffusion model: {self.tune_diffusion_model}")
+        # Check if any parameters are still trainable. If not, print a warning.
+        if not tune_projector and not tune_diffusion_model:
+            for name, p in self.named_parameters():
+                if p.requires_grad:
+                    print(f"Action head trainable parameter: {name}")
+        if not any(p.requires_grad for p in self.parameters()):
+            print("Warning: No action head trainable parameters found.")
+
+    def set_frozen_modules_to_eval_mode(self):
+        """
+        Huggingface will call model.train() at each training_step. To ensure
+        the expected behaviors for modules like dropout, batchnorm, etc., we
+        need to call model.eval() for the frozen modules.
+        """
+        if self.training:
+            if not self.tune_projector:
+                self.state_encoder.eval()
+                self.action_encoder.eval()
+                self.action_decoder.eval()
+                if self.config.add_pos_embed:
+                    self.position_embedding.eval()
+            if not self.tune_diffusion_model:
+                self.model.eval()
+
+    def sample_time(self, batch_size, device, dtype):
+        sample = self.beta_dist.sample([batch_size]).to(device, dtype=dtype)
+        return (self.config.noise_s - sample) / self.config.noise_s
+
+    def prepare_input(self, batch: dict) -> BatchFeature:
+        return BatchFeature(data=batch)
+
+    def process_backbone_output(self, backbone_output: BatchFeature) -> BatchFeature:
+        backbone_features = backbone_output["backbone_features"]
+        backbone_features = self.vlln(backbone_features)
+        backbone_features = self.vl_self_attention(backbone_features)
+        backbone_output["backbone_features"] = backbone_features
+        return backbone_output
+
+    def forward(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
+        # Set frozen modules to eval
+        self.set_frozen_modules_to_eval_mode()
+
+        backbone_output = self.process_backbone_output(backbone_output)
+
+        if self.config.expand_batch is not None:
+            for k, v in backbone_output.items():
+                ndim = len(v.shape)
+                factors = [self.config.expand_batch]
+                while len(factors) < ndim:
+                    factors.append(1)
+                factors = tuple(factors)
+                expanded = v.repeat(*factors)
+                backbone_output[k] = expanded
+
+            for k, v in action_input.items():
+                ndim = len(v.shape)
+                factors = [self.config.expand_batch]
+                while len(factors) < ndim:
+                    factors.append(1)
+                factors = tuple(factors)
+                expanded = v.repeat(*factors)
+                action_input[k] = expanded
+
+        # Get vision and language embeddings.
+        vl_embs = backbone_output.backbone_features
+        device = vl_embs.device
+
+        # Get embodiment ID.
+        embodiment_id = action_input.embodiment_id
+
+        # Embed state.
+        state_features = self.state_encoder(action_input.state, embodiment_id)
+
+        # Embed noised action trajectory.
+        actions = action_input.action
+        noise = torch.randn(actions.shape, device=actions.device, dtype=actions.dtype)
+        t = self.sample_time(actions.shape[0], device=actions.device, dtype=actions.dtype)
+        t = t[:, None, None]  # shape (B,1,1) for broadcast
+
+        noisy_trajectory = (1 - t) * noise + t * actions
+        velocity = actions - noise
+
+        # Convert (continuous) t -> discrete if needed
+        t_discretized = (t[:, 0, 0] * self.num_timestep_buckets).long()
+        action_features = self.action_encoder(noisy_trajectory, t_discretized, embodiment_id)
+
+        # Maybe add position embedding.
+        if self.config.add_pos_embed:
+            pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
+            pos_embs = self.position_embedding(pos_ids).unsqueeze(0)
+            action_features = action_features + pos_embs
+
+        # Join vision, language, state and action embedding along sequence dimension.
+        future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1)
+        sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1)
+
+        vl_attn_mask = backbone_output.backbone_attention_mask
+
+        model_output = self.model(
+            hidden_states=sa_embs,
+            encoder_hidden_states=vl_embs,
+            encoder_attention_mask=vl_attn_mask,
+            timestep=t_discretized,
+            return_all_hidden_states=False,  # NOTE (YL): not using flare now
+        )
+        pred = self.action_decoder(model_output, embodiment_id)
+        pred_actions = pred[:, -actions.shape[1] :]
+
+        # Slice out only the action portion of pred and target.
+        action_mask = action_input.action_mask
+        loss = F.mse_loss(pred_actions, velocity, reduction="none") * action_mask
+        loss = loss.sum() / action_mask.sum()
+        output_dict = {
+            "loss": loss,
+        }
+        return BatchFeature(data=output_dict)
+
+    @torch.no_grad()
+    def get_action(self, backbone_output: BatchFeature, action_input: BatchFeature) -> BatchFeature:
+        backbone_output = self.process_backbone_output(backbone_output)
+
+        # Get vision and language embeddings.
+        vl_embs = backbone_output.backbone_features
+        embodiment_id = action_input.embodiment_id
+
+        # Embed state.
+        state_features = self.state_encoder(action_input.state, embodiment_id)
+
+        # Set initial actions as the sampled noise.
+        batch_size = vl_embs.shape[0]
+        device = vl_embs.device
+        actions = torch.randn(
+            size=(batch_size, self.config.action_horizon, self.config.action_dim),
+            dtype=vl_embs.dtype,
+            device=device,
+        )
+
+        num_steps = self.num_inference_timesteps
+        dt = 1.0 / num_steps
+
+        # Run denoising steps.
+        for t in range(num_steps):
+            t_cont = t / float(num_steps)  # e.g. goes 0, 1/N, 2/N, ...
+            t_discretized = int(t_cont * self.num_timestep_buckets)
+
+            # Embed noised action trajectory.
+            timesteps_tensor = torch.full(size=(batch_size,), fill_value=t_discretized, device=device)
+            action_features = self.action_encoder(actions, timesteps_tensor, embodiment_id)
+            # Maybe add position embedding.
+            if self.config.add_pos_embed:
+                pos_ids = torch.arange(action_features.shape[1], dtype=torch.long, device=device)
+                pos_embs = self.position_embedding(pos_ids).unsqueeze(0)
+                action_features = action_features + pos_embs
+
+            # Join vision, language, state and action embedding along sequence dimension.
+            future_tokens = self.future_tokens.weight.unsqueeze(0).expand(vl_embs.shape[0], -1, -1)
+            sa_embs = torch.cat((state_features, future_tokens, action_features), dim=1)
+
+            # Run model forward.
+            model_output = self.model(
+                hidden_states=sa_embs,
+                encoder_hidden_states=vl_embs,
+                timestep=timesteps_tensor,
+            )
+            pred = self.action_decoder(model_output, embodiment_id)
+
+            pred_velocity = pred[:, -self.action_horizon :]
+
+            # Update actions using euler integration.
+            actions = actions + dt * pred_velocity
+        return BatchFeature(data={"action_pred": actions})
+
+    @property
+    def device(self):
+        return next(iter(self.parameters())).device
+
+    @property
+    def dtype(self):
+        return next(iter(self.parameters())).dtype

src/lerobot/policies/groot/configuration_groot.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python
+
+# Copyright 2024 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+from lerobot.optim.optimizers import AdamWConfig
+from lerobot.optim.schedulers import CosineDecayWithWarmupSchedulerConfig
+
+
+@PreTrainedConfig.register_subclass("groot")
+@dataclass
+class GrootConfig(PreTrainedConfig):
+    """Configuration for Groot policy wrapper."""
+
+    # Basic policy settings
+    n_obs_steps: int = 1
+    chunk_size: int = 50
+    n_action_steps: int = 50
+
+    # Dimension settings (must match pretrained GR00T model expectations)
+    # Maximum state dimension. Shorter states will be zero-padded.
+    max_state_dim: int = 64
+
+    # Maximum action dimension. Shorter actions will be zero-padded.
+    max_action_dim: int = 32
+
+    # Normalization (start with identity, adjust as needed)
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Image preprocessing (adjust to match Groot's expected input)
+    image_size: tuple[int, int] = (224, 224)
+
+    # Groot-specific model parameters (from groot_finetune_script.py)
+
+    # Path or HuggingFace model ID for the base Groot model
+    base_model_path: str = "nvidia/GR00T-N1.5-3B"
+
+    # HF repo ID (or local path) that hosts vocab.json and merges.txt for Eagle tokenizer.
+    tokenizer_assets_repo: str = "lerobot/eagle2hg-processor-groot-n1p5"
+
+    # Embodiment tag to use for training (e.g. 'new_embodiment', 'gr1')
+    embodiment_tag: str = "new_embodiment"
+
+    # Fine-tuning control arguments
+
+    # Whether to fine-tune the llm backbone
+    tune_llm: bool = False
+
+    # Whether to fine-tune the vision tower
+    tune_visual: bool = False
+
+    # Whether to fine-tune the projector
+    tune_projector: bool = True
+
+    # Whether to fine-tune the diffusion model
+    tune_diffusion_model: bool = True
+
+    # LoRA parameters (from groot_finetune_script.py)
+    # Rank for the LORA model. If 0, no LORA will be used.
+    lora_rank: int = 0
+
+    # Alpha value for the LORA model
+    lora_alpha: int = 16
+
+    # Dropout rate for the LORA model
+    lora_dropout: float = 0.1
+
+    # Whether to use the full model for LORA
+    lora_full_model: bool = False
+
+    # Training parameters (matching groot_finetune_script.py)
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.95, 0.999)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-5
+    warmup_ratio: float = 0.05
+    use_bf16: bool = True
+
+    # Dataset parameters
+    # Video backend to use for training ('decord' or 'torchvision_av')
+    video_backend: str = "decord"
+
+    # Whether to balance dataset weights in mixture datasets
+    balance_dataset_weights: bool = True
+
+    # Whether to sample trajectories weighted by their length
+    balance_trajectory_weights: bool = True
+
+    # Optional dataset paths for delegating training to Isaac-GR00T runner
+    dataset_paths: list[str] | None = None
+    output_dir: str = "./tmp/gr00t"
+    save_steps: int = 1000
+    max_steps: int = 10000
+    batch_size: int = 32
+    dataloader_num_workers: int = 8
+    report_to: str = "wandb"
+    resume: bool = False
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"n_action_steps ({self.n_action_steps}) cannot exceed chunk_size ({self.chunk_size})"
+            )
+
+        # groot_repo_path is now optional since we ported the components
+        # No validation needed
+
+    def validate_features(self) -> None:
+        """Validate and set up input/output features for Groot."""
+        image_features = [key for key, feat in self.input_features.items() if feat.type == FeatureType.VISUAL]
+        if not image_features:
+            raise ValueError(
+                "Groot policy requires at least one visual input feature. "
+                "No features of type FeatureType.VISUAL found in input_features."
+            )
+
+        if "observation.state" not in self.input_features:
+            state_feature = PolicyFeature(
+                type=FeatureType.STATE,
+                shape=(self.max_state_dim,),
+            )
+            self.input_features["observation.state"] = state_feature
+        else:
+            state_shape = self.input_features["observation.state"].shape
+            state_dim = state_shape[0] if state_shape else 0
+            if state_dim > self.max_state_dim:
+                raise ValueError(
+                    f"State dimension {state_dim} exceeds max_state_dim {self.max_state_dim}. "
+                    f"Either reduce state dimension or increase max_state_dim in config."
+                )
+
+        if "action" not in self.output_features:
+            action_feature = PolicyFeature(
+                type=FeatureType.ACTION,
+                shape=(self.max_action_dim,),
+            )
+            self.output_features["action"] = action_feature
+        else:
+            action_shape = self.output_features["action"].shape
+            action_dim = action_shape[0] if action_shape else 0
+            if action_dim > self.max_action_dim:
+                raise ValueError(
+                    f"Action dimension {action_dim} exceeds max_action_dim {self.max_action_dim}. "
+                    f"Either reduce action dimension or increase max_action_dim in config."
+                )
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        """Return optimizer configuration."""
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+        )
+
+    def get_scheduler_preset(self) -> CosineDecayWithWarmupSchedulerConfig:
+        """Return scheduler configuration."""
+        return CosineDecayWithWarmupSchedulerConfig(
+            num_warmup_steps=int(10000 * self.warmup_ratio),  # 5% warmup by default
+            num_decay_steps=10000,  # Adjust based on training steps
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.optimizer_lr * 0.1,
+        )
+
+    @property
+    def observation_delta_indices(self) -> None:
+        """Return indices for delta observations (None for Groot)."""
+        return None
+
+    @property
+    def action_delta_indices(self) -> list[int]:
+        """Return indices for delta actions."""
+        return list(range(min(self.chunk_size, 16)))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        """Return indices for delta rewards (None for Groot)."""
+        return None

src/lerobot/policies/groot/eagle2_hg_model/configuration_eagle2_5_vl.py

@@ -0,0 +1,135 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import copy
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.models.llama.configuration_llama import LlamaConfig
+from transformers.models.qwen2.configuration_qwen2 import Qwen2Config
+from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
+from transformers.models.siglip.configuration_siglip import SiglipVisionConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class Eagle25VLConfig(PretrainedConfig):
+    model_type = "eagle_2_5_vl"
+    is_composition = True
+    sub_configs = {"vision_config": SiglipVisionConfig, "text_config": Qwen2Config}
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        use_backbone_lora=0,
+        use_llm_lora=0,
+        pad2square=False,
+        select_layer=-4,
+        force_image_size=None,
+        downsample_ratio=0.5,
+        template=None,
+        dynamic_image_size=False,
+        use_thumbnail=False,
+        loss_version="v1",
+        min_dynamic_tiles=1,
+        max_dynamic_tiles=6,
+        mlp_checkpoint=False,
+        initializer_range=0.02,
+        _attn_implementation="flash_attention_2",
+        _attn_implementation_autoset=False,
+        llm_config=None,
+        image_token_index=None,
+        use_pixel_shuffle=True,
+        mlp_connector_layers=2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {"model_type": "siglip_vision_model"}
+            logger.info("vision_config is None. Initializing the InternVisionConfig with default values.")
+
+        if text_config is None:
+            text_config = {"architectures": ["Qwen2ForCausalLM"]}
+            logger.info(
+                "text_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`)."
+            )
+
+        if vision_config["model_type"] == "siglip_vision_model":
+            self.vision_config = SiglipVisionConfig(**vision_config)
+        else:
+            raise ValueError("Unsupported model_type: {}".format(vision_config["model_type"]))
+
+        if text_config["architectures"][0] == "LlamaForCausalLM":
+            self.text_config = LlamaConfig(**text_config)
+        elif text_config["architectures"][0] == "Qwen2ForCausalLM":
+            self.text_config = Qwen2Config(**text_config)
+        elif text_config["architectures"][0] == "Qwen3ForCausalLM":
+            self.text_config = Qwen3Config(**text_config)
+        else:
+            raise ValueError("Unsupported architecture: {}".format(text_config["architectures"][0]))
+        self.use_backbone_lora = use_backbone_lora
+        self.use_llm_lora = use_llm_lora
+        self.mlp_checkpoint = mlp_checkpoint
+        self.pad2square = pad2square
+        self.select_layer = select_layer
+        self.force_image_size = force_image_size
+        self.downsample_ratio = downsample_ratio
+        self.template = template
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.loss_version = loss_version
+        self.initializer_range = initializer_range
+        self.min_dynamic_tiles = min_dynamic_tiles
+        self.max_dynamic_tiles = max_dynamic_tiles
+        self.tie_word_embeddings = self.text_config.tie_word_embeddings
+        self._attn_implementation = _attn_implementation
+        self._attn_implementation_autoset = _attn_implementation_autoset
+        self.image_token_index = image_token_index
+        self.use_pixel_shuffle = use_pixel_shuffle
+        self.mlp_connector_layers = mlp_connector_layers
+        logger.info(f"min_dynamic_tiles: {self.min_dynamic_tiles}")
+        logger.info(f"max_dynamic_tiles: {self.max_dynamic_tiles}")
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["vision_config"] = self.vision_config.to_dict()
+        output["text_config"] = self.text_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        output["use_backbone_lora"] = self.use_backbone_lora
+        output["use_llm_lora"] = self.use_llm_lora
+        output["pad2square"] = self.pad2square
+        output["select_layer"] = self.select_layer
+        output["force_image_size"] = self.force_image_size
+        output["downsample_ratio"] = self.downsample_ratio
+        output["template"] = self.template
+        output["dynamic_image_size"] = self.dynamic_image_size
+        output["use_thumbnail"] = self.use_thumbnail
+        output["min_dynamic_tiles"] = self.min_dynamic_tiles
+        output["max_dynamic_tiles"] = self.max_dynamic_tiles
+        output["tie_word_embeddings"] = self.tie_word_embeddings
+        output["_attn_implementation"] = self._attn_implementation
+        output["_attn_implementation_autoset"] = self._attn_implementation_autoset
+        output["use_pixel_shuffle"] = self.use_pixel_shuffle
+        output["mlp_connector_layers"] = self.mlp_connector_layers
+        return output

src/lerobot/policies/groot/eagle2_hg_model/image_processing_eagle2_5_vl_fast.py

@@ -0,0 +1,504 @@
+# --------------------------------------------------------
+# NVIDIA
+# Copyright (c) 2025 NVIDIA
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+
+
+# copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py
+from typing import Optional
+
+from transformers.image_processing_utils import (
+    BatchFeature,
+    get_patch_output_size,
+)
+from transformers.image_processing_utils_fast import (
+    BaseImageProcessorFast,
+    DefaultFastImageProcessorKwargs,
+    group_images_by_shape,
+    reorder_images,
+)
+from transformers.image_utils import (
+    IMAGENET_STANDARD_MEAN,  # 0.5, 0.5, 0.5
+    IMAGENET_STANDARD_STD,  # 0.5, 0.5, 0.5
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+    get_image_size,
+    make_flat_list_of_images,
+    validate_kwargs,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    TensorType,
+    add_start_docstrings,
+    is_torch_available,
+    is_torchvision_v2_available,
+)
+from transformers.video_utils import VideoInput
+
+if is_torch_available():
+    import torch
+if is_torchvision_v2_available():
+    from torchvision.transforms.v2 import functional as F  # noqa: N812
+    from transformers.image_utils import pil_torch_interpolation_mapping
+else:
+    from torchvision.transforms import functional as F  # noqa: N812
+
+
+def crop(img: torch.Tensor, left: int, top: int, right: int, bottom: int) -> torch.Tensor:
+    """Crop the given numpy array.
+
+    Args:
+        img (torch.Tensor): Image to be cropped. Format should be (C, H, W).
+        left (int): The left coordinate of the crop box.
+        top (int): The top coordinate of the crop box.
+        right (int): The right coordinate of the crop box.
+        bottom (int): The bottom coordinate of the crop box.
+
+    Returns:
+        torch.Tensor: Cropped image.
+    """
+    if not isinstance(img, torch.Tensor):
+        raise TypeError(f"img should be torch.Tensor. Got {type(img)}")
+
+    if img.ndim not in [2, 3]:
+        raise ValueError(f"Image should have 2 or 3 dimensions. Got {img.ndim}")
+
+    img_height = img.shape[1]
+    img_width = img.shape[2]
+    if top < 0 or left < 0 or bottom > img_height or right > img_width:
+        raise ValueError("Crop coordinates out of bounds")
+
+    if top >= bottom or left >= right:
+        raise ValueError("Invalid crop coordinates")
+
+    return img[:, top:bottom, left:right]
+
+
+class Eagle25VLFastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    max_dynamic_tiles: int | None
+    min_dynamic_tiles: int | None
+    use_thumbnail: bool | None
+    pad_during_tiling: bool | None
+    do_pad: bool | None
+
+
+@add_start_docstrings(
+    "Constructs a fast ConvNeXT image processor. Based on [`SiglipImageProcessor`] with incorporation of processing each video frame.",
+    # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING, TODO: this was depreciated from transformers remove!
+    """
+        image_grid_pinpoints (`List[List[int]]`, *optional*):
+            A list of possible resolutions to use for processing high resolution images. The best resolution is selected
+            based on the original size of the image. Can be overridden by `image_grid_pinpoints` in the `preprocess`
+            method. Not used for processing videos.
+        do_pad (`bool`, *optional*):
+            Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+            number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+    """,
+)
+class Eagle25VLImageProcessorFast(BaseImageProcessorFast):
+    resample = PILImageResampling.BICUBIC
+    image_mean = IMAGENET_STANDARD_MEAN
+    image_std = IMAGENET_STANDARD_STD
+    size = {"height": 448, "width": 448}
+    default_to_square = False
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+    do_pad = True
+    max_dynamic_tiles = 12
+    min_dynamic_tiles = 1
+    use_thumbnail = True
+    pad_during_tiling = False
+    valid_kwargs = Eagle25VLFastImageProcessorKwargs
+    model_input_names = ["pixel_values_videos"]
+
+    def __init__(self, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @add_start_docstrings(
+        # BASE_IMAGE_PROCESSOR_FAST_DOCSTRING_PREPROCESS, TODO: this was depreciated from transformers remove!
+        """
+            max_dynamic_tiles (`int`, *optional*):
+                The maximum number of dynamic tiles to use for processing high resolution images.
+            min_dynamic_tiles (`int`, *optional*):
+                The minimum number of dynamic tiles to use for processing high resolution images.
+            use_thumbnail (`bool`, *optional*):
+                Whether to use a thumbnail for processing high resolution images.
+            pad_during_tiling (`bool`, *optional*):
+                Whether to pad the image during tiling.
+            do_pad (`bool`, *optional*):
+                    Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+                    number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+        """,
+    )
+
+    # NOTE(YL): we will overload the preprocess method to add the image_flags
+    # def preprocess(
+    #     self, images: ImageInput, **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs]
+    # ) -> BatchFeature:
+    #     return super().preprocess(images, **kwargs)
+
+    def _prepare_images_structure(
+        self,
+        images: ImageInput,
+        expected_ndims: int = 3,
+    ) -> ImageInput:
+        """
+        Prepare the images structure for processing.
+
+        Args:
+            images (`ImageInput`):
+                The input images to process.
+            expected_ndims (`int`, *optional*, defaults to 3):
+                Expected number of dimensions for the images (added for transformers >=4.53.0 compatibility).
+
+        Returns:
+            `ImageInput`: The images with a valid nesting.
+        """
+        return make_flat_list_of_images(images)
+
+    def _resize_for_patching(
+        self,
+        image: "torch.Tensor",
+        target_resolution: tuple,
+        interpolation: "F.InterpolationMode",
+        input_data_format: ChannelDimension,
+    ) -> "torch.Tensor":
+        """
+        Resizes an image to a target resolution while maintaining aspect ratio.
+
+        Args:
+            image ("torch.Tensor"):
+                The input image.
+            target_resolution (tuple):
+                The target resolution (height, width) of the image.
+            interpolation (`InterpolationMode`):
+                Resampling filter to use if resizing the image.
+            input_data_format (`ChannelDimension` or `str`):
+                The channel dimension format of the input image.
+
+        Returns:
+            "torch.Tensor": The resized and padded image.
+        """
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
+
+        # Resize the image
+        resized_image = F.resize(image, (new_height, new_width), interpolation=interpolation)
+
+        return resized_image
+
+    def find_closest_aspect_ratio(self, aspect_ratio, target_ratios, width, height, image_size):
+        """
+        previous version mainly focus on ratio.
+        We also consider area ratio here.
+        """
+        best_factor = float("-inf")
+        best_ratio = (1, 1)
+        area = width * height
+        for ratio in target_ratios:
+            target_aspect_ratio = ratio[0] / ratio[1]
+            # ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+            # area_ratio = (ratio[0] * ratio[1] * image_size * image_size) / area
+            """
+            new area > 60% of original image area is enough.
+            """
+            factor_based_on_area_n_ratio = min(
+                (ratio[0] * ratio[1] * image_size * image_size) / area, 0.6
+            ) * min(target_aspect_ratio / aspect_ratio, aspect_ratio / target_aspect_ratio)
+
+            if factor_based_on_area_n_ratio > best_factor:
+                best_factor = factor_based_on_area_n_ratio
+                best_ratio = ratio
+
+        return best_ratio
+
+    def _pad_for_patching(
+        self, image: "torch.Tensor", target_resolution: tuple, input_data_format: ChannelDimension
+    ) -> "torch.Tensor":
+        """
+        Pad an image to a target resolution while maintaining aspect ratio.
+        """
+        target_height, target_width = target_resolution
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
+
+        paste_x = (target_width - new_width) // 2
+        paste_y = (target_height - new_height) // 2
+
+        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x, paste_y])
+
+        return padded_image
+
+    def _get_image_patches(
+        self,
+        image: "torch.Tensor",
+        min_num: int,
+        max_num: int,
+        size: tuple,
+        tile_size: int,
+        use_thumbnail: bool,
+        interpolation: "F.InterpolationMode",
+        pad_during_tiling: bool,
+    ) -> list["torch.Tensor"]:
+        image_size = get_image_size(image, channel_dim=ChannelDimension.FIRST)
+        orig_height, orig_width = image_size
+        aspect_ratio = orig_width / orig_height
+
+        # calculate the existing image aspect ratio
+        target_ratios = {
+            (i, j)
+            for n in range(min_num, max_num + 1)
+            for i in range(1, n + 1)
+            for j in range(1, n + 1)
+            if i * j <= max_num and i * j >= min_num
+        }
+        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+        # find the closest aspect ratio to the target
+        target_aspect_ratio = self.find_closest_aspect_ratio(
+            aspect_ratio, target_ratios, orig_width, orig_height, tile_size
+        )
+
+        # calculate the target width and height
+        target_width = tile_size * target_aspect_ratio[0]
+        target_height = tile_size * target_aspect_ratio[1]
+        blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+        if pad_during_tiling:
+            resized_image = self._resize_for_patching(
+                image,
+                (target_height, target_width),
+                interpolation=interpolation,
+                input_data_format=ChannelDimension.FIRST,
+            )
+            padded_image = self._pad_for_patching(
+                resized_image,
+                (target_height, target_width),
+                input_data_format=ChannelDimension.FIRST,
+            )
+            image_used_to_split = padded_image
+        else:
+            image_used_to_split = F.resize(image, (target_height, target_width), interpolation=interpolation)
+
+        processed_tiles = []
+        for i in range(blocks):
+            box = (
+                (i % (target_width // tile_size)) * tile_size,
+                (i // (target_width // tile_size)) * tile_size,
+                ((i % (target_width // tile_size)) + 1) * tile_size,
+                ((i // (target_width // tile_size)) + 1) * tile_size,
+            )
+            # split the image
+            split_img = crop(image_used_to_split, box[0], box[1], box[2], box[3])
+            processed_tiles.append(split_img)
+        assert len(processed_tiles) == blocks
+
+        if use_thumbnail and len(processed_tiles) != 1:
+            thumbnail_img = F.resize(image, (tile_size, tile_size), interpolation=interpolation)
+            processed_tiles.append(thumbnail_img)
+
+        return processed_tiles
+
+    def _pad_for_batching(
+        self,
+        pixel_values: list["torch.Tensor"],
+    ) -> list["torch.Tensor"]:
+        """
+        Pads images on the `num_of_patches` dimension with zeros to form a batch of same number of patches.
+
+        Args:
+            pixel_values (`List[torch.Tensor]`):
+                An array of pixel values of each images of shape (`batch_size`, `num_patches`, `image_in_3D`)
+
+        Returns:
+            List[`torch.Tensor`]: The padded images.
+        """
+        max_patch = max(len(x) for x in pixel_values)
+        pixel_values = [
+            torch.nn.functional.pad(image, pad=[0, 0, 0, 0, 0, 0, 0, max_patch - image.shape[0]])
+            for image in pixel_values
+        ]
+
+        return pixel_values
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        max_dynamic_tiles: int,
+        min_dynamic_tiles: int,
+        use_thumbnail: bool,
+        pad_during_tiling: bool,
+        interpolation: Optional["F.InterpolationMode"],
+        do_center_crop: bool,
+        crop_size: SizeDict,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: float | list[float] | None,
+        image_std: float | list[float] | None,
+        do_pad: bool,
+        return_tensors: str | TensorType | None,
+        pad_size: SizeDict | None = None,  # Added for transformers >=4.53.0 compatibility
+        disable_grouping: bool | None = None,  # Added for transformers >=4.53.0 compatibility
+    ) -> BatchFeature:
+        processed_images = []
+        image_sizes = []
+        # Determine the size tuple
+        if size and size.height and size.width:
+            size_tuple = (size.height, size.width)
+        else:
+            size_tuple = (size.shortest_edge, size.shortest_edge)
+
+        # Determine the patch size
+        if crop_size and crop_size.height:
+            tile_size = crop_size.height
+        elif size and size.height:
+            tile_size = size.height
+        else:
+            tile_size = size.shortest_edge
+
+        for image in images:
+            image_patches = self._get_image_patches(
+                image,
+                min_num=min_dynamic_tiles,
+                max_num=max_dynamic_tiles,
+                size=size_tuple,
+                tile_size=tile_size,
+                use_thumbnail=use_thumbnail,
+                interpolation=interpolation,
+                pad_during_tiling=pad_during_tiling,
+            )
+
+            # Group images by size for batched processing
+            processed_image_patches_grouped = {}
+            # Added for transformers >=4.53.0 compatibility
+            grouped_image_patches, grouped_image_patches_index = group_images_by_shape(
+                image_patches,
+                disable_grouping=disable_grouping,
+            )
+
+            for shape, stacked_image_patches in grouped_image_patches.items():
+                if do_resize:
+                    stacked_image_patches = self.resize(
+                        image=stacked_image_patches,
+                        size=size,
+                        interpolation=interpolation,
+                    )
+                if do_center_crop:
+                    stacked_image_patches = self.center_crop(stacked_image_patches, crop_size)
+                # Fused rescale and normalize
+                stacked_image_patches = self.rescale_and_normalize(
+                    stacked_image_patches,
+                    do_rescale,
+                    rescale_factor,
+                    do_normalize,
+                    image_mean,
+                    image_std,
+                )
+                processed_image_patches_grouped[shape] = stacked_image_patches
+            processed_image_patches = reorder_images(
+                processed_image_patches_grouped, grouped_image_patches_index
+            )
+            processed_image_patches = (
+                torch.stack(processed_image_patches, dim=0) if return_tensors else processed_image_patches
+            )
+            processed_images.append(processed_image_patches)
+            image_sizes.append(get_image_size(image, ChannelDimension.FIRST))
+
+        if do_pad:
+            processed_images = self._pad_for_batching(processed_images)
+
+        # processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+        processed_images = torch.cat(processed_images, dim=0) if return_tensors else processed_images
+        return BatchFeature(
+            data={"pixel_values": processed_images, "image_sizes": image_sizes},
+            tensor_type=return_tensors,
+        )
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        videos: VideoInput = None,
+        **kwargs: Unpack[Eagle25VLFastImageProcessorKwargs],
+    ) -> BatchFeature:
+        validate_kwargs(
+            captured_kwargs=kwargs.keys(),
+            valid_processor_keys=self.valid_kwargs.__annotations__.keys(),
+        )
+        # Set default kwargs from self. This ensures that if a kwarg is not provided
+        # by the user, it gets its default value from the instance, or is set to None.
+        for kwarg_name in self.valid_kwargs.__annotations__:
+            kwargs.setdefault(kwarg_name, getattr(self, kwarg_name, None))
+
+        # Extract parameters that are only used for preparing the input images
+        do_convert_rgb = kwargs.pop("do_convert_rgb")
+        input_data_format = kwargs.pop("input_data_format")
+        device = kwargs.pop("device")
+        # Prepare input images
+        # transformers >= 4.53.0: uses _prepare_image_like_inputs instead of _prepare_input_images
+        if images is not None:
+            images = self._prepare_image_like_inputs(
+                images=images,
+                do_convert_rgb=do_convert_rgb,
+                input_data_format=input_data_format,
+                device=device,
+            )
+
+        if videos is not None:
+            videos = self._prepare_image_like_inputs(
+                images=videos,
+                do_convert_rgb=do_convert_rgb,
+                input_data_format=input_data_format,
+                device=device,
+            )
+
+        # Update kwargs that need further processing before being validated
+        kwargs = self._further_process_kwargs(**kwargs)
+
+        # Validate kwargs
+        self._validate_preprocess_kwargs(**kwargs)
+
+        # torch resize uses interpolation instead of resample
+        # Added for transformers >=4.53.0 compatibility
+        resample = kwargs.pop("resample", self.resample)
+        kwargs["interpolation"] = (
+            pil_torch_interpolation_mapping[resample]
+            if isinstance(resample, PILImageResampling | int)
+            else resample
+        )
+
+        # Filter kwargs to only include those accepted by _preprocess
+        valid_preprocess_kwargs = {
+            "do_resize",
+            "size",
+            "max_dynamic_tiles",
+            "min_dynamic_tiles",
+            "use_thumbnail",
+            "pad_during_tiling",
+            "interpolation",
+            "do_center_crop",
+            "crop_size",
+            "do_rescale",
+            "rescale_factor",
+            "do_normalize",
+            "image_mean",
+            "image_std",
+            "do_pad",
+            "return_tensors",
+            "pad_size",
+            "disable_grouping",
+        }
+        filtered_kwargs = {k: v for k, v in kwargs.items() if k in valid_preprocess_kwargs}
+        if images is not None:
+            return self._preprocess(images, **filtered_kwargs)
+        elif videos is not None:
+            return self._preprocess(videos, **filtered_kwargs)
+
+
+__all__ = ["Eagle25VLImageProcessorFast"]

src/lerobot/policies/groot/eagle2_hg_model/modeling_eagle2_5_vl.py

@@ -0,0 +1,395 @@
+# --------------------------------------------------------
+# NVIDIA
+# Copyright (c) 2025 NVIDIA
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+
+import inspect
+
+import torch
+import torch.utils.checkpoint as cp
+from peft import LoraConfig, get_peft_model
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers import GenerationConfig
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.llama.modeling_llama import LlamaForCausalLM
+from transformers.models.qwen2.modeling_qwen2 import Qwen2ForCausalLM
+from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
+from transformers.models.siglip.modeling_siglip import SiglipVisionModel
+from transformers.utils import add_start_docstrings, logging
+
+from .configuration_eagle2_5_vl import Eagle25VLConfig
+
+logger = logging.get_logger(__name__)
+
+
+# copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/modeling_llava_onevision.py#L241C1-L280C1
+EAGLE2_5_VL_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Eagle25VLConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Eagle2_5_VL Model outputting raw hidden-states without any specific head on top.",
+    EAGLE2_5_VL_START_DOCSTRING,
+)
+class Eagle25VLPreTrainedModel(PreTrainedModel):
+    config_class = Eagle25VLConfig
+    base_model_prefix = "model"
+    main_input_name = "input_ids"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "Qwen2DecoderLayer",
+        "LlamaDecoderLayer",
+        "Siglip2EncoderLayer",
+        "SiglipEncoderLayer",
+    ]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_cache_class = True
+    _supports_static_cache = True
+    _supports_quantized_cache = True
+    _supports_sdpa = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear | nn.Conv2d):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+class Eagle25VLForConditionalGeneration(Eagle25VLPreTrainedModel, GenerationMixin):
+    config_class = Eagle25VLConfig
+
+    def __init__(self, config: Eagle25VLConfig, vision_model=None, language_model=None):
+        super().__init__(config)
+
+        image_size = config.force_image_size or config.vision_config.image_size
+        patch_size = config.vision_config.patch_size
+        self.patch_size = patch_size
+        if config.use_pixel_shuffle:
+            self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio**2))
+        else:
+            self.num_image_token = int((image_size // patch_size) ** 2)
+
+        self.select_layer = config.select_layer
+        self.downsample_ratio = config.downsample_ratio
+        self.loss_version = config.loss_version
+        self.mlp_checkpoint = config.mlp_checkpoint
+        self.use_pixel_shuffle = config.use_pixel_shuffle
+        self.mlp_connector_layers = config.mlp_connector_layers
+        logger.info(f"num_image_token: {self.num_image_token}")
+        logger.info(f"mlp_checkpoint: {self.mlp_checkpoint}")
+        if vision_model is not None:
+            self.vision_model = vision_model
+        else:
+            if config.vision_config.model_type == "siglip_vision_model":
+                config.vision_config._attn_implementation = "flash_attention_2"
+                self.vision_model = SiglipVisionModel(config.vision_config)
+            else:
+                raise NotImplementedError(f"{config.vision_config.model_type} is not implemented.")
+
+        if language_model is not None:
+            self.language_model = language_model
+        else:
+            if config.text_config.architectures[0] == "LlamaForCausalLM":
+                self.language_model = LlamaForCausalLM(config.text_config)
+            elif config.text_config.architectures[0] == "Phi3ForCausalLM":
+                raise NotImplementedError("Phi3 is not implemented.")
+                # self.language_model = Phi3ForCausalLM(config.text_config)
+            elif config.text_config.architectures[0] == "Qwen2ForCausalLM":
+                assert config.text_config._attn_implementation == "flash_attention_2", (
+                    f"Qwen2 must use flash_attention_2 but got {config.text_config._attn_implementation}"
+                )
+                self.language_model = Qwen2ForCausalLM(config.text_config)
+            elif config.text_config.architectures[0] == "Qwen3ForCausalLM":
+                self.language_model = Qwen3ForCausalLM(config.text_config)
+            else:
+                raise NotImplementedError(f"{config.text_config.architectures[0]} is not implemented.")
+
+        vit_hidden_size = config.vision_config.hidden_size
+        llm_hidden_size = config.text_config.hidden_size
+
+        if config.mlp_connector_layers == 2:
+            self.mlp1 = nn.Sequential(
+                nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
+                nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
+                nn.GELU(),
+                nn.Linear(llm_hidden_size, llm_hidden_size),
+            )
+        elif config.mlp_connector_layers == 1 and config.use_pixel_shuffle:
+            self.mlp1 = nn.Sequential(
+                nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
+            )
+        elif config.mlp_connector_layers == 1 and not config.use_pixel_shuffle:
+            self.mlp1 = nn.Sequential(
+                nn.Linear(vit_hidden_size, llm_hidden_size),
+            )
+        else:
+            raise NotImplementedError(f"{config.mlp_connector_layers} is not implemented.")
+
+        self.image_token_index = config.image_token_index
+        self.neftune_alpha = None
+
+        if config.use_backbone_lora:
+            self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora)
+
+        self.use_llm_lora = config.use_llm_lora
+        if config.use_llm_lora:
+            self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora)
+
+        self.check_forward_kwargs()
+
+    def check_forward_kwargs(self):
+        # We intentionally avoid using **kwargs in forward because Hugging Face Transformers
+        # has special handling for functions with **kwargs parameters that would affect
+        # how our model is processed during training and inference.
+        forward_params = inspect.signature(self.forward).parameters
+        assert not any(k.kind == inspect.Parameter.VAR_KEYWORD for k in forward_params.values())
+
+    def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
+        lora_config = LoraConfig(
+            r=r,
+            target_modules=[
+                "self_attn.q_proj",
+                "self_attn.k_proj",
+                "self_attn.v_proj",
+                "self_attn.out_proj",
+                "mlp.fc1",
+                "mlp.fc2",
+            ],
+            lora_alpha=lora_alpha,
+            lora_dropout=lora_dropout,
+        )
+        self.vision_model = get_peft_model(self.vision_model, lora_config)
+        self.vision_model.print_trainable_parameters()
+
+    def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
+        lora_config = LoraConfig(
+            r=r,
+            target_modules=[
+                "self_attn.q_proj",
+                "self_attn.k_proj",
+                "self_attn.v_proj",
+                "self_attn.o_proj",
+                "mlp.gate_proj",
+                "mlp.down_proj",
+                "mlp.up_proj",
+            ],
+            lora_alpha=lora_alpha,
+            lora_dropout=lora_dropout,
+            task_type="CAUSAL_LM",
+        )
+        self.language_model = get_peft_model(self.language_model, lora_config)
+        self.language_model.enable_input_require_grads()
+        self.language_model.print_trainable_parameters()
+        self.use_llm_lora = True
+
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        image_flags: torch.LongTensor | None = None,
+        past_key_values: list[torch.FloatTensor] | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        num_tiles_list: list[torch.Tensor] | None = None,
+    ) -> tuple | CausalLMOutputWithPast:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        input_embeds = self.language_model.get_input_embeddings()(input_ids)
+
+        vit_embeds = self.extract_feature(pixel_values)
+
+        if image_flags is not None:
+            image_flags = image_flags.view(-1)
+            vit_embeds = vit_embeds[image_flags == 1]
+
+        b, n, c = input_embeds.shape
+        input_embeds = input_embeds.reshape(b * n, c)
+
+        input_ids = input_ids.reshape(b * n)
+        selected = input_ids == self.image_token_index
+        try:
+            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, c)
+        except Exception as e:
+            vit_embeds = vit_embeds.reshape(-1, c)
+            print(
+                f"warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, "
+                f"vit_embeds.shape={vit_embeds.shape}"
+            )
+            n_token = selected.sum()
+            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
+
+        input_embeds = input_embeds.reshape(b, n, c)
+
+        outputs = self.language_model(
+            inputs_embeds=input_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        logits = outputs.logits
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def pixel_shuffle(self, x, scale_factor=0.5):
+        n, w, h, c = x.size()
+        # N, W, H, C --> N, W, H * scale, C // scale
+        x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
+        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
+        x = x.permute(0, 2, 1, 3).contiguous()
+        # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
+        x = x.view(n, int(h * scale_factor), int(w * scale_factor), int(c / (scale_factor * scale_factor)))
+
+        x = x.permute(0, 2, 1, 3).contiguous()
+        return x
+
+    def extract_feature(self, pixel_values):
+        if self.select_layer == -1:
+            vit_embeds = self.vision_model(
+                pixel_values=pixel_values, output_hidden_states=False, return_dict=True
+            )
+            if hasattr(vit_embeds, "last_hidden_state"):
+                vit_embeds = vit_embeds.last_hidden_state
+
+        else:
+            vit_embeds = self.vision_model(
+                pixel_values=pixel_values, output_hidden_states=True, return_dict=True
+            ).hidden_states[self.select_layer]
+
+        if self.use_pixel_shuffle:
+            h = w = int(vit_embeds.shape[1] ** 0.5)
+            vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
+            vit_embeds = self.pixel_shuffle(
+                vit_embeds, scale_factor=self.downsample_ratio
+            )  # torch.Size([B, 1024, 1024]) -> torch.Size([B, 16, 16, 4096])
+            vit_embeds = vit_embeds.reshape(
+                vit_embeds.shape[0], -1, vit_embeds.shape[-1]
+            )  # torch.Size([B, 16, 16, 4096]) -> torch.Size([B, 256, 4096])
+
+        if self.mlp_checkpoint and vit_embeds.requires_grad:
+            vit_embeds = cp.checkpoint(self.mlp1, vit_embeds)
+        else:
+            vit_embeds = self.mlp1(vit_embeds)
+
+        return vit_embeds
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor | None = None,
+        input_ids: torch.FloatTensor | None = None,
+        attention_mask: torch.LongTensor | None = None,
+        visual_features: torch.FloatTensor | None = None,
+        generation_config: GenerationConfig | None = None,
+        output_hidden_states: bool | None = None,
+        image_sizes: list[tuple[int, int]] | None = None,
+        **generate_kwargs,
+    ) -> torch.LongTensor:
+        if pixel_values is not None:
+            if visual_features is not None:
+                vit_embeds = visual_features
+            else:
+                vit_embeds = self.extract_feature(pixel_values)
+
+            input_embeds = self.language_model.get_input_embeddings()(input_ids)
+            b, n, c = input_embeds.shape
+            input_embeds = input_embeds.reshape(b * n, c)
+
+            input_ids = input_ids.reshape(b * n)
+            selected = input_ids == self.config.image_token_index
+            assert selected.sum() != 0
+            input_embeds[selected] = vit_embeds.reshape(-1, c).to(input_embeds.device)
+
+            input_embeds = input_embeds.reshape(b, n, c)
+        else:
+            input_embeds = self.language_model.get_input_embeddings()(input_ids)
+
+        if "use_cache" not in generate_kwargs:
+            generate_kwargs["use_cache"] = True
+
+        outputs = self.language_model.generate(
+            inputs_embeds=input_embeds,
+            attention_mask=attention_mask,
+            generation_config=generation_config,
+            output_hidden_states=output_hidden_states,
+            **generate_kwargs,
+        )
+
+        return outputs
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_input_embeddings
+    def set_input_embeddings(self, value):
+        self.language_model.set_input_embeddings(value)
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.language_model.get_output_embeddings()
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.set_decoder
+    def set_decoder(self, decoder):
+        self.language_model.set_decoder(decoder)
+
+    # Copied from transformers.models.llava_next.modeling_llava_next.LlavaNextForConditionalGeneration.get_decoder
+    def get_decoder(self):
+        return self.language_model.get_decoder()

src/lerobot/policies/groot/eagle2_hg_model/processing_eagle2_5_vl.py

@@ -0,0 +1,518 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for Eagle25VL.
+copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/processing_llava_onevision.py
+"""
+
+import base64
+import os
+import re
+from io import BytesIO
+
+import requests
+import torch
+from PIL import Image
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+from transformers.utils import logging
+from transformers.video_utils import VideoInput
+
+logger = logging.get_logger(__name__)
+
+
+FRAME_FACTOR = 2
+FPS = 2.0
+FPS_MIN_FRAMES = 4
+FPS_MAX_FRAMES = 256
+
+
+def to_rgb(pil_image: Image.Image) -> Image.Image:
+    if pil_image.mode == "RGBA":
+        white_background = Image.new("RGB", pil_image.size, (255, 255, 255))
+        white_background.paste(pil_image, mask=pil_image.split()[3])  # Use alpha channel as mask
+        return white_background
+    else:
+        return pil_image.convert("RGB")
+
+
+def fetch_image(ele: dict[str, str | Image.Image]) -> Image.Image:
+    image = ele["image"] if "image" in ele else ele["image_url"]
+    image_obj = None
+    if isinstance(image, Image.Image):
+        image_obj = image
+    elif image.startswith("http://") or image.startswith("https://"):
+        response = requests.get(image, stream=True, timeout=10)
+        image_obj = Image.open(BytesIO(response.content))
+    elif image.startswith("file://"):
+        image_obj = Image.open(image[7:])
+    elif image.startswith("data:image"):
+        if "base64," in image:
+            _, base64_data = image.split("base64,", 1)
+            data = base64.b64decode(base64_data)
+            image_obj = Image.open(BytesIO(data))
+    else:
+        image_obj = Image.open(image)
+    if image_obj is None:
+        raise ValueError(
+            f"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}"
+        )
+    image = to_rgb(image_obj)
+    if "scale_factor" in ele:
+        scale_factor = ele["scale_factor"]
+        image = image.resize((image.width * scale_factor, image.height * scale_factor), Image.BILINEAR)
+    return image
+
+
+class Eagle25VLProcessorKwargs(ProcessingKwargs, total=False):
+    # see processing_utils.ProcessingKwargs documentation for usage.
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+        "images_kwargs": {},
+        "videos_kwargs": {"max_dynamic_tiles": 1},
+    }
+
+
+class Eagle25VLProcessor(ProcessorMixin):
+    r"""
+    Constructs a Eagle25VL processor which wraps a Eagle25VL video processor, Eagle25VL image processor and a Eagle25VL tokenizer into a single processor.
+
+    [`Eagle25VLProcessor`] offers all the functionalities of [`Eagle25VLVideoProcessor`], [`Eagle25VLImageProcessor`] and [`Eagle25VLTokenizer`]. See the
+    [`~Eagle25VLVideoProcessor.__call__`], [`~Eagle25VLProcessor.__call__`] and [`~Eagle25VLProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`LlavaOnevisionImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        num_image_tokens (`int`, *optional*):
+            Number of image tokens for one imagethat will be returned by vision tower.
+        vision_feature_select_strategy (`str`, *optional*):
+            The feature selection strategy used to select the vision feature from the vision backbone.
+            Should be same as in model's config
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+        image_token (`str`, *optional*, defaults to `"<image>"`):
+            Special token used to denote image location.
+        video_token (`str`, *optional*, defaults to `"<video>"`):
+            Special token used to denote video location.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = [
+        "chat_template",
+        "num_image_tokens",
+        "vision_feature_select_strategy",
+        "image_token",
+        "video_token",
+        "images_kwargs",
+        "videos_kwargs",
+        "text_kwargs",
+    ]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        vision_feature_select_strategy=None,
+        chat_template=None,
+        image_token="<IMG_CONTEXT>",  # nosec: B107
+        video_token="<IMG_CONTEXT>",  # nosec: B107
+        tokens_per_tile=256,
+        image_placeholder="image",
+        video_placeholder="video",
+        image_start_token="<img>",
+        image_end_token="</img>",
+        **kwargs,
+    ):
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.image_token = tokenizer.image_token if hasattr(tokenizer, "image_token") else image_token
+        self.video_token = tokenizer.video_token if hasattr(tokenizer, "video_token") else video_token
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+        self.video_token_id = (
+            tokenizer.video_token_id
+            if getattr(tokenizer, "video_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.video_token)
+        )
+        self.image_placeholder = image_placeholder
+        self.video_placeholder = video_placeholder
+        self.tokens_per_tile = tokens_per_tile
+        self.image_start_token = image_start_token
+        self.image_end_token = image_end_token
+        if "auto_map" in kwargs:
+            self.auto_map = kwargs["auto_map"]
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def replace_media_placeholder(
+        self, text, image_list, video_list, timestamps_list, fps_list, **output_kwargs
+    ):
+        num_of_images_in_this_sample = 0
+        num_of_videos_in_this_sample = 0
+        # Regular expression pattern to match formats like <image-1> or <video-2>
+        pattern = re.compile(rf"<({self.image_placeholder}|{self.video_placeholder})-(\d+)>")
+        unified_frame_list = []
+
+        # image_min_dynamic_tiles = output_kwargs["images_kwargs"].get(
+        #     "min_dynamic_tiles", self.image_processor.min_dynamic_tiles
+        # )
+        # image_max_dynamic_tiles = output_kwargs["images_kwargs"].get(
+        #     "max_dynamic_tiles", self.image_processor.max_dynamic_tiles
+        # )
+        # image_use_thumbnail = output_kwargs["images_kwargs"].get(
+        #     "use_thumbnail", self.image_processor.use_thumbnail
+        # )
+        video_min_dynamic_tiles = output_kwargs["videos_kwargs"].get(
+            "min_dynamic_tiles", self.image_processor.min_dynamic_tiles
+        )
+        video_max_dynamic_tiles = output_kwargs["videos_kwargs"].get(
+            "max_dynamic_tiles", self.image_processor.max_dynamic_tiles
+        )
+        video_use_thumbnail = output_kwargs["videos_kwargs"].get(
+            "use_thumbnail", self.image_processor.use_thumbnail
+        )
+
+        tile_size = self.image_processor.size.get("height", 448)
+
+        # Function to replace tags in a single text
+        def replace_in_text(text):
+            # repl callback function for each match replacement operation
+            def repl(match):
+                nonlocal unified_frame_list
+                nonlocal num_of_images_in_this_sample
+                nonlocal num_of_videos_in_this_sample
+                media_type = match.group(1)  # 'image' or 'video'
+                idx_in_list = int(match.group(2)) - 1  # Convert to list index (0-based)
+                # Select the corresponding path based on media type
+                idx_mapper = {
+                    0: "first",
+                    1: "second",
+                    2: "third",
+                    3: "fourth",
+                    4: "fifth",
+                    5: "sixth",
+                    6: "seventh",
+                    7: "eighth",
+                    8: "ninth",
+                    9: "tenth",
+                }
+                if media_type == "image":
+                    image_inputs = self.image_processor(
+                        images=[image_list[idx_in_list]],
+                        videos=None,
+                        **output_kwargs["images_kwargs"],
+                    )
+                    num_all_tiles = image_inputs["pixel_values"].shape[0]
+                    special_placeholder = f"<image {idx_in_list + 1}>{self.image_start_token}{self.image_token * num_all_tiles * self.tokens_per_tile}{self.image_end_token}"
+                    unified_frame_list.append(image_inputs)
+                    num_of_images_in_this_sample += 1
+
+                elif media_type == "video":
+                    video_inputs = self.image_processor(
+                        images=None,
+                        videos=[video_list[idx_in_list]],
+                        **output_kwargs["videos_kwargs"],
+                    )
+                    num_all_tiles = video_inputs["pixel_values"].shape[0]
+                    image_sizes = video_inputs["image_sizes"]
+                    if timestamps_list is not None and -1 not in timestamps_list:
+                        frame_timestamps = timestamps_list[idx_in_list]
+                    else:
+                        frame_timestamps = None
+                    sampled_fps = fps_list[idx_in_list] if fps_list is not None else None
+
+                    num_of_tiles_each_frame = [
+                        self.get_number_tiles_based_on_image_size(
+                            image_size,
+                            video_min_dynamic_tiles,
+                            video_max_dynamic_tiles,
+                            video_use_thumbnail,
+                            tile_size,
+                        )
+                        for image_size in image_sizes
+                    ]
+                    assert sum(num_of_tiles_each_frame) == num_all_tiles, (
+                        f"The number of tiles in each frame is not equal to the total number of tiles: {sum(num_of_tiles_each_frame)} != {num_all_tiles}"
+                    )
+
+                    if frame_timestamps is not None:
+                        assert len(frame_timestamps) == len(num_of_tiles_each_frame), (
+                            f"The number of timestamps is not equal to the number of frames: {len(frame_timestamps)} != {len(num_of_tiles_each_frame)}"
+                        )
+                        special_placeholder = [
+                            f"Frame {i + 1} sample at {frame_timestamps[i]:.2f}s: {self.image_start_token}{self.image_token * num_of_tiles * self.tokens_per_tile}{self.image_end_token}"
+                            for i, num_of_tiles in enumerate(num_of_tiles_each_frame)
+                        ]
+                    else:
+                        special_placeholder = [
+                            f"Frame {i + 1}: {self.image_start_token}{self.image_token * num_of_tiles * self.tokens_per_tile}{self.image_end_token}"
+                            for i, num_of_tiles in enumerate(num_of_tiles_each_frame)
+                        ]
+
+                    if sampled_fps is not None:
+                        special_placeholder = (
+                            f"The {idx_mapper[idx_in_list]} video sampled with {sampled_fps:.2f} fps: "
+                            + "".join(special_placeholder)
+                        )
+                    else:
+                        special_placeholder = f"The {idx_mapper[idx_in_list]} video: " + "".join(
+                            special_placeholder
+                        )
+                    unified_frame_list.append(video_inputs)
+                    num_of_videos_in_this_sample += 1
+                else:
+                    raise ValueError(f"Unknown media type: {media_type}")
+                return special_placeholder
+
+            return pattern.sub(repl, text)
+
+        text = replace_in_text(text)
+        if len(unified_frame_list) > 0:
+            pixel_values = torch.cat([frame["pixel_values"] for frame in unified_frame_list])
+            image_sizes = torch.cat([frame["image_sizes"] for frame in unified_frame_list])
+        else:
+            pixel_values = None
+            image_sizes = None
+        return (
+            text,
+            pixel_values,
+            image_sizes,
+            num_of_images_in_this_sample,
+            num_of_videos_in_this_sample,
+        )
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] = None,
+        audio=None,
+        videos: VideoInput = None,
+        **kwargs: Unpack[Eagle25VLProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        LlavaNextImageProcessor's [`~LlavaNextImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
+        of the above two methods for more information.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of a video input to be fed to a model. Returned when `videos` is not `None`.
+            - **image_sizes** -- Size of each image that will be used to unpad an image. Returned when `images` is not `None`.
+        """
+
+        output_kwargs = self._merge_kwargs(
+            Eagle25VLProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        if isinstance(text, str):
+            text_list = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+        elif isinstance(text, list) and isinstance(text[0], str):
+            text_list = text
+
+        if images is None:
+            images = []
+        if videos is None:
+            videos = []
+
+        pixel_values_list = []
+        image_sizes_list = []
+        new_sample_list = []
+        image_start_idx = 0
+        video_start_idx = 0
+        timestamps_batch = output_kwargs["videos_kwargs"].pop("timestamps", None)
+        fps_batch = output_kwargs["videos_kwargs"].pop("fps", None)
+        for sample in text_list:
+            timestamps_list = timestamps_batch[video_start_idx:] if timestamps_batch is not None else None
+            fps_list = fps_batch[video_start_idx:] if fps_batch is not None else None
+            (
+                sample,
+                pixel_values,
+                image_sizes,
+                num_of_images_in_this_sample,
+                num_of_videos_in_this_sample,
+            ) = self.replace_media_placeholder(
+                sample,
+                images[image_start_idx:],
+                videos[video_start_idx:],
+                timestamps_list,
+                fps_list,
+                **output_kwargs,
+            )
+            new_sample_list.append(sample)
+            if pixel_values is not None:
+                pixel_values_list.append(pixel_values)
+                image_sizes_list.append(image_sizes)
+            image_start_idx += num_of_images_in_this_sample
+            video_start_idx += num_of_videos_in_this_sample
+
+        if len(pixel_values_list) > 0:
+            image_inputs = {
+                "pixel_values": torch.cat(pixel_values_list),
+                "image_sizes": torch.cat(image_sizes_list),
+            }
+        else:
+            image_inputs = {}
+        video_inputs = {}
+        text_inputs = self.tokenizer(new_sample_list, **output_kwargs["text_kwargs"])
+        return BatchFeature(data={**text_inputs, **image_inputs, **video_inputs})
+
+    def get_number_tiles_based_on_image_size(
+        self, image_size: tuple, min_num: int, max_num: int, use_thumbnail: bool, tile_size: int
+    ) -> int:
+        """
+        Get the number of tiles based on the image size.
+        """
+        orig_height, orig_width = image_size
+        aspect_ratio = orig_width / orig_height
+        # calculate the existing image aspect ratio
+        target_ratios = {
+            (i, j)
+            for n in range(min_num, max_num + 1)
+            for i in range(1, n + 1)
+            for j in range(1, n + 1)
+            if i * j <= max_num and i * j >= min_num
+        }
+        target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+        # find the closest aspect ratio to the target
+        target_aspect_ratio = self.image_processor.find_closest_aspect_ratio(
+            aspect_ratio, target_ratios, orig_width, orig_height, tile_size
+        )
+        tiles_num = target_aspect_ratio[0] * target_aspect_ratio[1]
+        if use_thumbnail and tiles_num > 1:
+            tiles_num += 1
+        return tiles_num
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+    # override to save video-config in a separate config file
+    def save_pretrained(self, save_directory, **kwargs):
+        if os.path.isfile(save_directory):
+            raise ValueError(f"Provided path ({save_directory}) should be a directory, not a file")
+        os.makedirs(save_directory, exist_ok=True)
+
+        outputs = super().save_pretrained(save_directory, **kwargs)
+        return outputs
+
+    # override to load video-config from a separate config file
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        processor = super().from_pretrained(pretrained_model_name_or_path, **kwargs)
+
+        # if return_unused_kwargs a tuple is returned where the second element is 'unused_kwargs'
+        if isinstance(processor, tuple):
+            processor = processor[0]
+        return processor
+
+    # Copy from https://github.com/QwenLM/Qwen2.5-VL/blob/main/qwen-vl-utils/src/qwen_vl_utils/vision_process.py
+    def process_vision_info(
+        self,
+        conversations: list[dict] | list[list[dict]],
+        return_video_kwargs: bool = False,
+    ) -> tuple[list[Image.Image] | None, list[torch.Tensor | list[Image.Image]] | None, dict | None]:
+        vision_infos = self.extract_vision_info(conversations)
+        ## Read images or videos
+        image_inputs = []
+        video_inputs = []
+        video_sample_fps_list = []
+        video_timestamps_list = []
+        for vision_info in vision_infos:
+            if "image" in vision_info or "image_url" in vision_info:
+                image_inputs.append(fetch_image(vision_info))
+            else:
+                raise ValueError("image, image_url or video should in content.")
+        if len(image_inputs) == 0:
+            image_inputs = None
+        if len(video_inputs) == 0:
+            video_inputs = None
+        if return_video_kwargs:
+            return (
+                image_inputs,
+                video_inputs,
+                {"fps": video_sample_fps_list, "timestamps": video_timestamps_list},
+            )
+        return image_inputs, video_inputs
+
+    def extract_vision_info(self, conversations: list[dict] | list[list[dict]]) -> list[dict]:
+        vision_infos = []
+        if isinstance(conversations[0], dict):
+            conversations = [conversations]
+        for conversation in conversations:
+            for message in conversation:
+                if isinstance(message["content"], list):
+                    for ele in message["content"]:
+                        if (
+                            "image" in ele
+                            or "image_url" in ele
+                            or "video" in ele
+                            or ele["type"] in ("image", "image_url", "video")
+                        ):
+                            vision_infos.append(ele)
+        return vision_infos
+
+
+__all__ = ["Eagle25VLProcessor"]

src/lerobot/policies/groot/groot_n1.py

@@ -0,0 +1,376 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import TYPE_CHECKING
+
+import numpy as np
+import torch
+import torch.nn as nn
+from huggingface_hub import snapshot_download
+from huggingface_hub.errors import HFValidationError, RepositoryNotFoundError
+
+from lerobot.utils.import_utils import _transformers_available
+
+# Conditional import for type checking and lazy loading
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoConfig, AutoModel, PretrainedConfig, PreTrainedModel
+    from transformers.feature_extraction_utils import BatchFeature
+else:
+    AutoConfig = None
+    AutoModel = None
+    PretrainedConfig = object
+    PreTrainedModel = object
+    BatchFeature = None
+
+try:
+    import tree
+except ImportError:
+    tree = None
+
+from lerobot.policies.groot.action_head.flow_matching_action_head import (
+    FlowmatchingActionHead,
+    FlowmatchingActionHeadConfig,
+)
+from lerobot.policies.groot.utils import ensure_eagle_cache_ready
+from lerobot.utils.constants import HF_LEROBOT_HOME
+
+DEFAULT_VENDOR_EAGLE_PATH = str((Path(__file__).resolve().parent / "eagle2_hg_model").resolve())
+DEFAULT_TOKENIZER_ASSETS_REPO = "lerobot/eagle2hg-processor-groot-n1p5"
+
+
+class EagleBackbone(nn.Module):
+    def __init__(
+        self,
+        tune_llm: bool = False,
+        tune_visual: bool = False,
+        select_layer: int = -1,
+        reproject_vision: bool = False,
+        use_flash_attention: bool = False,
+        load_bf16: bool = False,
+        eagle_path: str = DEFAULT_VENDOR_EAGLE_PATH,
+        tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS_REPO,
+        project_to_dim: int = 1536,
+    ):
+        """
+        Args:
+            tune_llm: whether to tune the LLM model (default: True)
+            tune_visual: whether to tune the visual model (default: False)
+        """
+        super().__init__()
+        assert not reproject_vision, "Reproject vision is not implemented here, set to False"
+
+        # Prefer loading Eagle model config from the cache directory where vendor files were copied.
+        vendor_dir = DEFAULT_VENDOR_EAGLE_PATH
+        cache_dir = HF_LEROBOT_HOME / tokenizer_assets_repo
+        try:
+            ensure_eagle_cache_ready(vendor_dir, cache_dir, tokenizer_assets_repo)
+        except Exception as exc:  # nosec: B110
+            print(f"[GROOT] Warning: failed to prepare Eagle cache for backbone: {exc}")
+
+        config = AutoConfig.from_pretrained(str(cache_dir), trust_remote_code=True)
+        self.eagle_model = AutoModel.from_config(config, trust_remote_code=True)
+
+        if project_to_dim is not None:
+            self.eagle_linear = torch.nn.Linear(2048, project_to_dim)
+        else:
+            self.eagle_linear = torch.nn.Identity()
+
+        # needed since we don't use these layers. Also saves compute
+        while len(self.eagle_model.language_model.model.layers) > select_layer:
+            self.eagle_model.language_model.model.layers.pop(-1)
+
+        self.select_layer = select_layer
+        self.set_trainable_parameters(tune_llm, tune_visual)
+
+    def set_trainable_parameters(self, tune_llm: bool, tune_visual: bool):
+        self.tune_llm = tune_llm
+        self.tune_visual = tune_visual
+        for p in self.parameters():
+            p.requires_grad = True
+        if not tune_llm:
+            self.eagle_model.language_model.requires_grad_(False)
+        if not tune_visual:
+            self.eagle_model.vision_model.requires_grad_(False)
+            self.eagle_model.mlp1.requires_grad_(False)
+        print(f"Tune backbone llm: {self.tune_llm}")
+        print(f"Tune backbone visual: {self.tune_visual}")
+        # Check if any parameters are still trainable. If not, print a warning.
+        if not tune_llm and not tune_visual:
+            for name, p in self.named_parameters():
+                if p.requires_grad:
+                    print(f"Backbone trainable parameter: {name}")
+        if not any(p.requires_grad for p in self.parameters()):
+            print("Warning: No backbone trainable parameters found.")
+
+    def set_frozen_modules_to_eval_mode(self):
+        """
+        Huggingface will call model.train() at each training_step. To ensure
+        the expected behaviors for modules like dropout, batchnorm, etc., we
+        need to call model.eval() for the frozen modules.
+        """
+        if self.training:
+            if self.eagle_model.language_model and not self.tune_llm:
+                self.eagle_model.language_model.eval()
+            if self.eagle_model.vision_model and not self.tune_visual:
+                self.eagle_model.vision_model.eval()
+
+    def prepare_input(self, batch: dict) -> BatchFeature:
+        return BatchFeature(data=batch)
+
+    def forward_eagle(self, vl_input: BatchFeature) -> BatchFeature:
+        eagle_prefix = "eagle_"
+        eagle_input = {
+            k.removeprefix(eagle_prefix): v for k, v in vl_input.items() if k.startswith(eagle_prefix)
+        }
+        del eagle_input["image_sizes"]
+
+        eagle_output = self.eagle_model(**eagle_input, output_hidden_states=True, return_dict=True)
+        eagle_features = eagle_output.hidden_states[self.select_layer]
+
+        eagle_features = self.eagle_linear(eagle_features)
+        return eagle_features, eagle_input["attention_mask"]
+
+    def forward(self, vl_input: BatchFeature) -> BatchFeature:
+        self.set_frozen_modules_to_eval_mode()
+
+        eagle_embeds, eagle_mask = self.forward_eagle(vl_input)
+
+        # YL (TODO HACK): to resolve DDP issue when tune_visual=True
+        # Ensure all trainable parameters in vision_model are used in the forward pass for DDP compatibility
+        if self.training and self.tune_visual:
+            dummy_term = torch.tensor(
+                0.0, device=eagle_embeds.device, dtype=eagle_embeds.dtype, requires_grad=True
+            )
+            for param in self.eagle_model.vision_model.parameters():
+                if param.requires_grad:
+                    dummy_term = dummy_term + 0.0 * param.sum()
+            eagle_embeds = eagle_embeds + dummy_term
+
+        return BatchFeature(
+            data={"backbone_features": eagle_embeds, "backbone_attention_mask": eagle_mask}
+        )  # [B, T2, hidden_size]
+
+
+BACKBONE_FEATURE_KEY = "backbone_features"
+ACTION_KEY = "action_pred"
+LOSS_KEY = "loss"
+ERROR_MSG = "Error: unexpected input/output"
+N_COLOR_CHANNELS = 3
+
+
+# config
+@dataclass
+class GR00TN15Config(PretrainedConfig):
+    model_type = "gr00t_n1_5"
+    backbone_cfg: dict = field(init=False, metadata={"help": "Backbone configuration."})
+
+    action_head_cfg: dict = field(init=False, metadata={"help": "Action head configuration."})
+
+    action_horizon: int = field(init=False, metadata={"help": "Action horizon."})
+
+    action_dim: int = field(init=False, metadata={"help": "Action dimension."})
+    compute_dtype: str = field(default="float32", metadata={"help": "Compute dtype."})
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        for key, value in kwargs.items():
+            setattr(self, key, value)
+
+
+# real model
+class GR00TN15(PreTrainedModel):
+    supports_gradient_checkpointing = True
+    config_class = GR00TN15Config
+    """
+    we expect the backbone output to have a key 'backbone_features' with shape (batch_size, n, hidden_size)
+    here n is variable and can be e.g. time, 1 or user specified
+    we expect the action head output to have a key 'action_pred' with shape (batch_size, time, action_dim) during inference time
+    we expect these to have type BatchFeature, and they can of course have many other user specified keys too
+    """
+
+    def __init__(
+        self,
+        config: GR00TN15Config,
+        local_model_path: str,
+    ):
+        assert isinstance(config.backbone_cfg, dict)
+        assert isinstance(config.action_head_cfg, dict)
+
+        super().__init__(config)
+        self.local_model_path = local_model_path
+
+        self.backbone = EagleBackbone(**config.backbone_cfg)
+        action_head_cfg = FlowmatchingActionHeadConfig(**config.action_head_cfg)
+        self.action_head = FlowmatchingActionHead(action_head_cfg)
+
+        self.action_horizon = config.action_horizon
+        self.action_dim = config.action_dim
+        self.compute_dtype = config.compute_dtype
+
+    def validate_inputs(self, inputs):
+        # NOTE -- this should be handled internally by the model
+        # however, doing that will likely be breaking changes -- so we'll need to do it after the deadline
+
+        detected_error = False
+        error_msg = ERROR_MSG
+        if "action" in inputs:
+            action = inputs["action"]
+            # In inference, action may be omitted or None; validate only when it's a tensor.
+            if action is None:
+                pass  # allow None during inference
+            elif isinstance(action, torch.Tensor):
+                shape_ok = (
+                    len(action.shape) == 3
+                    and action.shape[1] == self.action_horizon
+                    and action.shape[2] == self.action_dim
+                )
+                if not shape_ok:
+                    error_msg += f"\n{action.shape=}"
+                    detected_error = True
+            else:
+                # Unexpected non-tensor type provided for action
+                error_msg += f"\nInvalid type for action: {type(action)}"
+                detected_error = True
+
+        if "video" in inputs:
+            video = inputs["video"]
+            type_ok = isinstance(video, np.ndarray)
+            dtype_ok = video.dtype == np.uint8
+            shape_ok = len(video.shape) == 6 and video.shape[3] == N_COLOR_CHANNELS
+            if not type_ok:
+                error_msg += f"\n{type(video)=}"
+                detected_error = True
+            if not dtype_ok:
+                error_msg += f"\n{video.dtype=}"
+                detected_error = True
+            if not shape_ok:
+                error_msg += f"\n{video.shape=}"
+                detected_error = True
+
+        if detected_error:
+            raise ValueError(error_msg)
+
+    def validate_data(self, action_head_outputs, backbone_outputs, is_training):
+        fail_backbone = (
+            not isinstance(backbone_outputs, BatchFeature) or BACKBONE_FEATURE_KEY not in backbone_outputs
+        )
+
+        if fail_backbone:
+            error_msg = ERROR_MSG
+            error_msg += f"\n{isinstance(backbone_outputs, BatchFeature)=}"
+            error_msg += f"\n{BACKBONE_FEATURE_KEY in backbone_outputs=}"
+            error_msg += f"\n{backbone_outputs[BACKBONE_FEATURE_KEY].shape=}"
+            raise ValueError(error_msg)
+
+        fail_action_head = (not isinstance(action_head_outputs, BatchFeature)) or not (
+            (
+                LOSS_KEY in action_head_outputs and is_training
+            )  # there might not be an action prediction during training
+            or (
+                ACTION_KEY in action_head_outputs
+                and action_head_outputs[ACTION_KEY].shape[1] == self.action_horizon
+                and action_head_outputs[ACTION_KEY].shape[2] == self.action_dim
+            )
+        )
+
+        if fail_action_head:
+            error_msg = ERROR_MSG
+            error_msg += f"\n{isinstance(action_head_outputs, BatchFeature)=}"
+            error_msg += f"\n{LOSS_KEY in action_head_outputs=}"
+            error_msg += f"\n{action_head_outputs[ACTION_KEY].shape=}"
+            error_msg += f"\n{self.action_horizon=}"
+            error_msg += f"\n{self.action_dim=}"
+            raise ValueError(error_msg)
+
+    def forward(
+        self,
+        inputs: dict,
+    ) -> BatchFeature:
+        backbone_inputs, action_inputs = self.prepare_input(inputs)
+        backbone_outputs = self.backbone(backbone_inputs)
+        action_head_outputs = self.action_head(backbone_outputs, action_inputs)
+        self.validate_data(action_head_outputs, backbone_outputs, is_training=True)
+        return action_head_outputs
+
+    def get_action(
+        self,
+        inputs: dict,
+    ) -> BatchFeature:
+        backbone_inputs, action_inputs = self.prepare_input(inputs)
+        # Because the behavior of backbones remains the same for training and inference, we can use `forward` for backbones.
+        backbone_outputs = self.backbone(backbone_inputs)
+        action_head_outputs = self.action_head.get_action(backbone_outputs, action_inputs)
+        self.validate_data(action_head_outputs, backbone_outputs, is_training=False)
+        return action_head_outputs
+
+    def prepare_input(self, inputs) -> tuple[BatchFeature, BatchFeature]:
+        self.validate_inputs(inputs)
+        backbone_inputs = self.backbone.prepare_input(inputs)
+        action_inputs = self.action_head.prepare_input(inputs)
+
+        def to_device_with_maybe_dtype(x):
+            # Cast floating tensors to a memory-efficient compute dtype when requested.
+            # Rationale: Upcasting backbone activations to fp32 significantly increases VRAM.
+            # When compute_dtype is bfloat16, prefer bf16 for activations to match AMP behavior.
+            if not isinstance(x, torch.Tensor):
+                return x
+            if torch.is_floating_point(x):
+                if getattr(self, "compute_dtype", None) == "bfloat16":
+                    return x.to(self.device, dtype=torch.bfloat16)
+                # Fallback: preserve previous behavior if not using bf16 compute
+                return x.to(self.device, dtype=self.action_head.dtype)
+            # Non-floating tensors: move device only
+            return x.to(self.device)
+
+        backbone_inputs = tree.map_structure(to_device_with_maybe_dtype, backbone_inputs)
+        action_inputs = tree.map_structure(to_device_with_maybe_dtype, action_inputs)
+        return backbone_inputs, action_inputs
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
+        tune_visual = kwargs.pop("tune_visual", True)
+        tune_llm = kwargs.pop("tune_llm", False)
+        tune_projector = kwargs.pop("tune_projector", True)
+        tune_diffusion_model = kwargs.pop("tune_diffusion_model", True)
+
+        print(f"Loading pretrained dual brain from {pretrained_model_name_or_path}")
+        print(f"Tune backbone vision tower: {tune_visual}")
+        print(f"Tune backbone LLM: {tune_llm}")
+        print(f"Tune action head projector: {tune_projector}")
+        print(f"Tune action head DiT: {tune_diffusion_model}")
+
+        # get the current model path being downloaded
+        try:
+            # NOTE(YL) This downloads the model to the local cache and returns the local path to the model
+            # saved in ~/.cache/huggingface/hub/
+            local_model_path = snapshot_download(pretrained_model_name_or_path, repo_type="model")
+            # HFValidationError, RepositoryNotFoundError
+        except (HFValidationError, RepositoryNotFoundError):
+            print(
+                f"Model not found or avail in the huggingface hub. Loading from local path: {pretrained_model_name_or_path}"
+            )
+            local_model_path = pretrained_model_name_or_path
+
+        pretrained_model = super().from_pretrained(
+            local_model_path, local_model_path=local_model_path, **kwargs
+        )
+
+        pretrained_model.backbone.set_trainable_parameters(tune_visual=tune_visual, tune_llm=tune_llm)
+        pretrained_model.action_head.set_trainable_parameters(
+            tune_projector=tune_projector, tune_diffusion_model=tune_diffusion_model
+        )
+        return pretrained_model

src/lerobot/policies/groot/modeling_groot.py

@@ -0,0 +1,198 @@
+#!/usr/bin/env python
+
+# Copyright 2024 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Groot Policy Wrapper for LeRobot Integration
+
+Minimal integration that delegates to Isaac-GR00T components where possible
+without porting their code. The intent is to:
+
+- Download and load the pretrained GR00T model via GR00TN15.from_pretrained
+- Optionally align action horizon similar to gr00t_finetune.py
+- Expose predict_action via GR00T model.get_action
+- Provide a training forward that can call the GR00T model forward if batch
+  structure matches.
+
+Notes:
+- Dataset loading and full training orchestration is handled by Isaac-GR00T
+  TrainRunner in their codebase. If you want to invoke that flow end-to-end
+  from LeRobot, see `GrootPolicy.finetune_with_groot_runner` below.
+"""
+
+import os
+from collections import deque
+
+import torch
+from torch import Tensor
+
+from lerobot.policies.groot.configuration_groot import GrootConfig
+from lerobot.policies.groot.groot_n1 import GR00TN15
+from lerobot.policies.pretrained import PreTrainedPolicy
+
+
+class GrootPolicy(PreTrainedPolicy):
+    """Wrapper around external Groot model for LeRobot integration."""
+
+    name = "groot"
+    config_class = GrootConfig
+
+    def __init__(self, config: GrootConfig):
+        """Initialize Groot policy wrapper."""
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+
+        # Initialize GR00T model using ported components
+        self._groot_model = self._create_groot_model()
+
+        self.reset()
+
+    def _create_groot_model(self):
+        """Create and initialize the GR00T model using Isaac-GR00T API.
+
+        This is only called when creating a NEW policy (not when loading from checkpoint).
+
+        Steps (delegating to Isaac-GR00T):
+        1) Download and load pretrained model via GR00TN15.from_pretrained
+        2) Align action horizon with data_config if provided
+        """
+        # Handle Flash Attention compatibility issues
+        self._handle_flash_attention_compatibility()
+
+        model = GR00TN15.from_pretrained(
+            pretrained_model_name_or_path=self.config.base_model_path,
+            tune_llm=self.config.tune_llm,
+            tune_visual=self.config.tune_visual,
+            tune_projector=self.config.tune_projector,
+            tune_diffusion_model=self.config.tune_diffusion_model,
+        )
+
+        model.compute_dtype = "bfloat16" if self.config.use_bf16 else model.compute_dtype
+        model.config.compute_dtype = model.compute_dtype
+
+        return model
+
+    def reset(self):
+        """Reset policy state when environment resets."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
+        """Training forward pass.
+
+        Delegates to Isaac-GR00T model.forward when inputs are compatible.
+        """
+        # Build a clean input dict for GR00T: keep only tensors GR00T consumes
+        allowed_base = {"state", "state_mask", "action", "action_mask", "embodiment_id"}
+        groot_inputs = {
+            k: v
+            for k, v in batch.items()
+            if (k in allowed_base or k.startswith("eagle_")) and not (k.startswith("next.") or k == "info")
+        }
+
+        # Get device from model parameters
+        device = next(self.parameters()).device
+
+        # Run GR00T forward under bf16 autocast when enabled to reduce activation memory
+        # Rationale: Matches original GR00T finetuning (bf16 compute, fp32 params) and avoids fp32 upcasts.
+        with torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=self.config.use_bf16):
+            outputs = self._groot_model.forward(groot_inputs)
+
+        # Isaac-GR00T returns a BatchFeature; loss key is typically 'loss'
+        loss = outputs.get("loss")
+
+        loss_dict = {"loss": loss.item()}
+
+        return loss, loss_dict
+
+    @torch.no_grad()
+    def predict_action_chunk(self, batch: dict[str, Tensor]) -> Tensor:
+        """Predict a chunk of actions for inference by delegating to Isaac-GR00T.
+
+        Returns a tensor of shape (B, n_action_steps, action_dim).
+        """
+        self.eval()
+
+        # Build a clean input dict for GR00T: keep only tensors GR00T consumes
+        # Preprocessing is handled by the processor pipeline, so we just filter the batch
+        # NOTE: During inference, we should NOT pass action/action_mask (that's what we're predicting)
+        allowed_base = {"state", "state_mask", "embodiment_id"}
+        groot_inputs = {
+            k: v
+            for k, v in batch.items()
+            if (k in allowed_base or k.startswith("eagle_")) and not (k.startswith("next.") or k == "info")
+        }
+
+        # Get device from model parameters
+        device = next(self.parameters()).device
+
+        # Use bf16 autocast for inference to keep memory low and match backbone dtype
+        with torch.autocast(device_type=device.type, dtype=torch.bfloat16, enabled=self.config.use_bf16):
+            outputs = self._groot_model.get_action(groot_inputs)
+
+        actions = outputs.get("action_pred")
+
+        original_action_dim = self.config.output_features["action"].shape[0]
+        actions = actions[:, :, :original_action_dim]
+
+        return actions
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select single action from action queue."""
+        self.eval()
+
+        if len(self._action_queue) == 0:
+            actions = self.predict_action_chunk(batch)
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    # -------------------------
+    # Internal helpers
+    # -------------------------
+    def _handle_flash_attention_compatibility(self) -> None:
+        """Handle Flash Attention compatibility issues by setting environment variables.
+
+        This addresses the common 'undefined symbol' error that occurs when Flash Attention
+        is compiled against a different PyTorch version than what's currently installed.
+        """
+
+        # Set environment variables to handle Flash Attention compatibility
+        # These help with symbol resolution issues
+        os.environ.setdefault("FLASH_ATTENTION_FORCE_BUILD", "0")
+        os.environ.setdefault("FLASH_ATTENTION_SKIP_CUDA_BUILD", "0")
+
+        # Try to import flash_attn and handle failures gracefully
+        try:
+            import flash_attn
+
+            print(f"[GROOT] Flash Attention version: {flash_attn.__version__}")
+        except ImportError as e:
+            print(f"[GROOT] Flash Attention not available: {e}")
+            print("[GROOT] Will use fallback attention mechanism")
+        except Exception as e:
+            if "undefined symbol" in str(e):
+                print(f"[GROOT] Flash Attention compatibility issue detected: {e}")
+                print("[GROOT] This is likely due to PyTorch/Flash Attention version mismatch")
+                print("[GROOT] Consider reinstalling Flash Attention with compatible version:")
+                print("  pip uninstall flash-attn")
+                print("  pip install --no-build-isolation flash-attn==2.6.3")
+                print("[GROOT] Continuing with fallback attention mechanism")
+            else:
+                print(f"[GROOT] Flash Attention error: {e}")
+                print("[GROOT] Continuing with fallback attention mechanism")

src/lerobot/policies/groot/processor_groot.py

@@ -0,0 +1,664 @@
+#!/usr/bin/env python
+
+# Copyright 2024 NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+import torch
+from einops import rearrange
+from PIL import Image
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import AutoProcessor, ProcessorMixin
+else:
+    AutoProcessor = None
+    ProcessorMixin = object
+
+from lerobot.configs.types import (
+    FeatureType,
+    NormalizationMode,
+    PolicyFeature,
+)
+from lerobot.policies.groot.configuration_groot import GrootConfig
+from lerobot.processor import (
+    AddBatchDimensionProcessorStep,
+    DeviceProcessorStep,
+    PolicyAction,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    ProcessorStepRegistry,
+    RenameObservationsProcessorStep,
+)
+from lerobot.processor.converters import (
+    policy_action_to_transition,
+    transition_to_policy_action,
+)
+from lerobot.processor.core import EnvTransition, TransitionKey
+from lerobot.utils.constants import (
+    HF_LEROBOT_HOME,
+    POLICY_POSTPROCESSOR_DEFAULT_NAME,
+    POLICY_PREPROCESSOR_DEFAULT_NAME,
+)
+
+# Defaults for Eagle processor locations
+DEFAULT_TOKENIZER_ASSETS_REPO = "lerobot/eagle2hg-processor-groot-n1p5"
+
+
+def make_groot_pre_post_processors(
+    config: GrootConfig, dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Create preprocessor and postprocessor for Groot policy.
+
+    This creates a processing pipeline that transforms LeRobot data format into
+    the format expected by Isaac-GR00T models:
+
+    Preprocessing steps:
+    1. Optional key renaming (dataset-specific key mapping)
+    2. Add batch dimension to unbatched data
+    3. Pack video/state/action/language/embodiment and apply optional min-max normalization before padding
+    4. Encode video+language with Eagle VLM into intermediate eagle_content
+    5. Collate eagle_content into batched eagle_* tensors
+    6. Move tensors to device (GPU)
+
+    NOTE: We optionally apply min-max normalization to STATE and ACTION using
+    dataset-provided statistics prior to padding, mapping values to [-1, 1].
+    This mirrors SO100-style preprocessing and keeps scales consistent with GR00T.
+
+    Args:
+        config: Groot configuration containing data_config, embodiment_tag, etc.
+        dataset_stats: Optional per-key min/max statistics for normalization before padding.
+
+    Returns:
+        Tuple of (preprocessor, postprocessor) pipelines
+    """
+
+    # Get horizon/dimension parameters from config
+    # These should match the config used for the pretrained model
+    # Default values match most GR00T configs (state_horizon=1, action_horizon=16)
+    state_horizon = 1
+    # CRITICAL: Pretrained GR00T models use action_horizon=16 max!
+    # The model architecture hardcodes this limit
+    action_horizon = min(config.chunk_size, 16)
+    max_state_dim = config.max_state_dim
+    max_action_dim = config.max_action_dim
+
+    # Pass raw dataset_stats; normalization will occur inside pack step before padding
+    padded_stats = dataset_stats or {}
+
+    # Define feature specs for optional normalization steps
+    _features: dict[str, PolicyFeature] = {
+        # Observation features (only add those we may normalize)
+        "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(state_horizon, max_state_dim)),
+        # Action feature
+        "action": PolicyFeature(type=FeatureType.ACTION, shape=(action_horizon, max_action_dim)),
+    }
+
+    # Normalize STATE and ACTION with min_max (SO100-like default)
+    _norm_map = {
+        FeatureType.ACTION: NormalizationMode.MIN_MAX,
+        FeatureType.STATE: NormalizationMode.MIN_MAX,
+    }
+
+    # Determine env action dimension from config (simple, object-like PolicyFeature)
+    try:
+        env_action_dim = int(config.output_features["action"].shape[0])
+    except Exception:
+        env_action_dim = 0
+
+    input_steps: list[ProcessorStep] = [
+        # 1. Rename keys if needed (e.g., dataset-specific camera names)
+        # Leave empty for now - add mappings if your dataset uses different key names
+        RenameObservationsProcessorStep(rename_map={}),
+        # 2. Add batch dimension for single samples
+        AddBatchDimensionProcessorStep(),
+        # 3. Pack video/state/action/language/embodiment; apply optional min-max normalization before padding
+        GrootPackInputsStep(
+            state_horizon=state_horizon,
+            action_horizon=action_horizon,
+            max_state_dim=max_state_dim,
+            max_action_dim=max_action_dim,
+            language_key="task",
+            formalize_language=False,
+            embodiment_tag=config.embodiment_tag,
+            normalize_min_max=True,
+            stats=padded_stats,
+        ),
+        # 4. Eagle encode (creates eagle_content)
+        GrootEagleEncodeStep(
+            tokenizer_assets_repo=config.tokenizer_assets_repo,
+        ),
+        # 5. Collate eagle_content -> eagle_* tensors
+        GrootEagleCollateStep(
+            tokenizer_assets_repo=config.tokenizer_assets_repo,
+        ),
+        # 6. Move to device
+        DeviceProcessorStep(device=config.device),
+    ]
+
+    # Postprocessing: slice to env action dim and unnormalize to env scale, then move to CPU
+    output_steps: list[ProcessorStep] = [
+        GrootActionUnpackUnnormalizeStep(
+            env_action_dim=env_action_dim,
+            stats=padded_stats,
+            normalize_min_max=True,
+        ),
+        # Finally, move to CPU for env interaction
+        DeviceProcessorStep(device="cpu"),
+    ]
+
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+# GR00T specific processor steps
+
+
+def _to_uint8_np_bhwc(img_t: torch.Tensor) -> np.ndarray:
+    # img_t: (B, C, H, W) float in [0,1] or uint8
+    if img_t.dtype.is_floating_point:
+        img_t = (img_t.clamp(0, 1) * 255.0).to(torch.uint8)
+    return rearrange(img_t.cpu().numpy(), "b c h w -> b h w c")
+
+
+def _build_eagle_processor(tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS_REPO) -> ProcessorMixin:
+    # Validate that the cache directory is ready. If not, instruct the user.
+    cache_dir = HF_LEROBOT_HOME / tokenizer_assets_repo
+    required = [
+        cache_dir / "processor_config.json",
+        cache_dir / "preprocessor_config.json",
+        cache_dir / "image_processing_eagle2_5_vl_fast.py",
+    ]
+    if not all(p.exists() for p in required):
+        raise FileNotFoundError(
+            f"[GROOT] Eagle processor cache at '{cache_dir}' is not populated. "
+            "Vendor files are copied during model creation. Create the policy/model first, "
+            "or call ensure_eagle_cache_ready() before building processors."
+        )
+    proc = AutoProcessor.from_pretrained(str(cache_dir), trust_remote_code=True, use_fast=True)
+    proc.tokenizer.padding_side = "left"
+    return proc
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="groot_pack_inputs_v3")
+class GrootPackInputsStep(ProcessorStep):
+    state_horizon: int = 1
+    action_horizon: int = 16
+    max_state_dim: int = 64
+    max_action_dim: int = 32
+    language_key: str = "task"
+    formalize_language: bool = False
+    embodiment_tag: str = "new_embodiment"
+    embodiment_mapping: dict[str, int] = field(
+        default_factory=lambda: {
+            "new_embodiment": 31,  # Match original GR00T EMBODIMENT_TAG_MAPPING
+            "oxe_droid": 17,
+            "agibot_genie1": 26,
+            "gr1": 24,
+            "so100": 2,
+            "unitree_g1": 3,
+        }
+    )
+    # Min-max normalization (SO100-like) applied BEFORE padding
+    normalize_min_max: bool = True
+    stats: dict[str, dict[str, Any]] | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        obs = transition.get(TransitionKey.OBSERVATION, {}) or {}
+        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}) or {}
+
+        def _align_vec(vec: Any, target_dim: int, *, default: float) -> torch.Tensor:
+            t = torch.as_tensor(vec)
+            t = t.flatten().to(
+                dtype=torch.float32,
+                device=next(
+                    (v.device for v in obs.values() if isinstance(v, torch.Tensor)), torch.device("cpu")
+                ),
+            )
+            d = int(t.shape[-1]) if t.numel() > 0 else 0
+            if d == target_dim:
+                return t
+            if d < target_dim:
+                pad = torch.full((target_dim - d,), default, dtype=t.dtype, device=t.device)
+                return torch.cat([t, pad], dim=0)
+            return t[:target_dim]
+
+        def _min_max_norm(x: torch.Tensor, key: str) -> torch.Tensor:
+            if not self.normalize_min_max:
+                return x
+            if self.stats is None or key not in self.stats:
+                return x
+            stats_k = self.stats[key]
+            last_dim = x.shape[-1]
+            min_v = _align_vec(stats_k.get("min", torch.zeros(last_dim)), last_dim, default=0.0)
+            max_v = _align_vec(stats_k.get("max", torch.ones(last_dim)), last_dim, default=1.0)
+            denom = max_v - min_v
+            mask = denom != 0
+            safe_denom = torch.where(mask, denom, torch.ones_like(denom))
+            mapped = 2 * (x - min_v) / safe_denom - 1
+            return torch.where(mask, mapped, torch.zeros_like(mapped))
+
+        # 1) Video (B, T=1, V, H, W, C) uint8
+        img_keys = sorted([k for k in obs if k.startswith("observation.images.")])
+        if not img_keys and "observation.image" in obs:
+            img_keys = ["observation.image"]
+        if img_keys:
+            cams = [_to_uint8_np_bhwc(obs[k]) for k in img_keys]
+            video = np.stack(cams, axis=1)  # (B, V, H, W, C)
+            video = np.expand_dims(video, axis=1)  # (B, 1, V, H, W, C)
+            # GR00T validates that video.shape[3] == 3 (channels), so reorder to (B, T, V, C, H, W)
+            video = np.transpose(video, (0, 1, 2, 5, 3, 4))  # (B, 1, V, C, H, W)
+            obs["video"] = video
+            # Drop raw images to avoid confusion downstream
+            for k in img_keys:
+                obs.pop(k, None)
+
+        # 2) Language (string)
+        lang = comp.get(self.language_key)
+        if isinstance(lang, list):
+            lang = lang[0] if len(lang) > 0 else None
+        if not lang:
+            lang = "Perform the task."
+        if self.formalize_language:
+            lang = (lang or "").lower()
+            lang = "".join(ch for ch in lang if ch.isalnum() or ch.isspace())
+        comp["language"] = lang
+
+        # 3) State/state_mask -> (B, 1, max_state_dim)
+        if "observation.state" in obs:
+            state = obs["observation.state"]  # (B, D)
+            if state.dim() != 2:
+                raise ValueError(f"state must be (B, D), got {tuple(state.shape)}")
+            bsz, d = state.shape
+            # Normalize BEFORE padding
+            if self.normalize_min_max:
+                state = _min_max_norm(state, "observation.state")
+            state = state.unsqueeze(1)  # (B, 1, D)
+            if d > self.max_state_dim:
+                state = state[:, :, : self.max_state_dim]
+                d = self.max_state_dim
+            elif d < self.max_state_dim:
+                pad = torch.zeros(bsz, 1, self.max_state_dim - d, dtype=state.dtype, device=state.device)
+                state = torch.cat([state, pad], dim=2)
+            state_mask = torch.zeros(bsz, 1, self.max_state_dim, dtype=torch.bool, device=state.device)
+            state_mask[:, :, :d] = True
+            obs["state"] = state
+            obs["state_mask"] = state_mask
+
+        # 4) Action/action_mask -> (B, action_horizon, max_action_dim)
+        action = transition.get(TransitionKey.ACTION)
+        if isinstance(action, torch.Tensor):
+            # Normalize BEFORE temporal expansion/padding
+            if self.normalize_min_max:
+                if action.dim() == 2:
+                    action = _min_max_norm(action, "action")
+                elif action.dim() == 3:
+                    b, t, d = action.shape
+                    flat = action.reshape(b * t, d)
+                    flat = _min_max_norm(flat, "action")
+                    action = flat.view(b, t, d)
+            if action.dim() == 2:
+                action = action.unsqueeze(1).repeat(1, self.action_horizon, 1)
+            elif action.dim() == 3:
+                b, t, d = action.shape
+                if t < self.action_horizon:
+                    last = action[:, -1:, :]
+                    pad = last.repeat(1, self.action_horizon - t, 1)
+                    action = torch.cat([action, pad], dim=1)
+                elif t > self.action_horizon:
+                    action = action[:, : self.action_horizon, :]
+            else:
+                raise ValueError(f"action must be (B, D) or (B, T, D), got {tuple(action.shape)}")
+
+            b, t, d = action.shape
+            if d > self.max_action_dim:
+                action = action[:, :, : self.max_action_dim]
+                d = self.max_action_dim
+            elif d < self.max_action_dim:
+                pad = torch.zeros(b, t, self.max_action_dim - d, dtype=action.dtype, device=action.device)
+                action = torch.cat([action, pad], dim=2)
+            action_mask = torch.zeros(b, t, self.max_action_dim, dtype=torch.bool, device=action.device)
+            action_mask[:, :, :d] = True
+            transition[TransitionKey.ACTION] = action
+            comp["action_mask"] = action_mask
+
+        # 5) Embodiment id as LongTensor (B,)
+        emb_id = self.embodiment_mapping.get(self.embodiment_tag, 0)
+        # Infer batch size/device from any tensor in obs or action
+        bsz = None
+        device = torch.device("cpu")
+        for v in list(obs.values()) + [transition.get(TransitionKey.ACTION)]:
+            if isinstance(v, torch.Tensor):
+                bsz = v.shape[0]
+                device = v.device
+                break
+        if bsz is None and "video" in obs and isinstance(obs["video"], np.ndarray):
+            bsz = obs["video"].shape[0]
+        if bsz is None:
+            bsz = 1
+        comp["embodiment_id"] = torch.full((bsz,), emb_id, dtype=torch.long, device=device)
+
+        transition[TransitionKey.OBSERVATION] = obs
+        transition[TransitionKey.COMPLEMENTARY_DATA] = comp
+        return transition
+
+    # Pipeline API requirement: declare how features change (we keep it simple)
+    def transform_features(self, features):
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns a serializable dictionary of the processor's configuration.
+
+        Excludes 'stats' since they are saved separately via state_dict().
+        """
+        return {
+            "state_horizon": self.state_horizon,
+            "action_horizon": self.action_horizon,
+            "max_state_dim": self.max_state_dim,
+            "max_action_dim": self.max_action_dim,
+            "language_key": self.language_key,
+            "formalize_language": self.formalize_language,
+            "embodiment_tag": self.embodiment_tag,
+            "embodiment_mapping": self.embodiment_mapping,
+            "normalize_min_max": self.normalize_min_max,
+        }
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """
+        Returns normalization statistics as a flat state dictionary.
+
+        This enables saving stats to safetensors files, similar to normalizer_processor.
+        """
+        if not self.stats:
+            return {}
+
+        flat: dict[str, torch.Tensor] = {}
+        for key, sub in self.stats.items():
+            for stat_name, value in sub.items():
+                tensor = torch.as_tensor(value).cpu()
+                flat[f"{key}.{stat_name}"] = tensor
+        return flat
+
+    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+        """
+        Loads normalization statistics from a flat state dictionary.
+
+        This enables loading stats from safetensors files during from_pretrained.
+        """
+        if not state:
+            return
+
+        reconstructed: dict[str, dict[str, Any]] = {}
+        for flat_key, tensor in state.items():
+            if "." in flat_key:
+                key, stat_name = flat_key.rsplit(".", 1)
+                if key not in reconstructed:
+                    reconstructed[key] = {}
+                reconstructed[key][stat_name] = tensor
+
+        if reconstructed:
+            self.stats = reconstructed
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="groot_eagle_encode_v3")
+class GrootEagleEncodeStep(ProcessorStep):
+    tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS_REPO
+    _proc: ProcessorMixin | None = field(default=None, init=False, repr=False)
+
+    @property
+    def proc(self) -> ProcessorMixin:
+        if self._proc is None:
+            self._proc = _build_eagle_processor(self.tokenizer_assets_repo)
+        return self._proc
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        obs = transition.get(TransitionKey.OBSERVATION, {}) or {}
+        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}) or {}
+
+        if "video" not in obs:
+            return transition
+
+        video = obs["video"]  # (B, T, V, H, W, C) uint8
+        lang = comp.get("language", "Perform the task.")
+        if isinstance(lang, list):
+            lang = lang[0] if len(lang) > 0 else "Perform the task."
+
+        bsz = video.shape[0]
+        eagle_contents: list[dict[str, Any]] = []
+        for b in range(bsz):
+            vt = video[b]  # (T, V, C, H, W) after reorder
+            if vt.ndim != 5:
+                # Fallback: assume (T, V, H, W, C)
+                t, v, h, w, c = vt.shape
+                flat = rearrange(vt, "t v h w c -> (t v) h w c")
+            else:
+                t, v, c, h, w = vt.shape
+                flat = rearrange(vt, "t v c h w -> (t v) h w c")
+            images = [Image.fromarray(flat[i]) for i in range(t * v)]
+            # Format language as string list representation to match Original GROOT
+            lang_formatted = str([lang])
+            text_content = [{"type": "text", "text": lang_formatted}]
+            image_content = [{"type": "image", "image": img} for img in images]
+            conv = [{"role": "user", "content": image_content + text_content}]
+            text_list = [self.proc.apply_chat_template(conv, tokenize=False, add_generation_prompt=True)]
+            img_inputs, vid_inputs = self.proc.process_vision_info(conv)
+            eagle_contents.append(
+                {
+                    "text_list": text_list,
+                    "image_inputs": img_inputs,
+                    "video_inputs": vid_inputs,
+                }
+            )
+
+        comp["eagle_content"] = eagle_contents
+        transition[TransitionKey.OBSERVATION] = obs
+        transition[TransitionKey.COMPLEMENTARY_DATA] = comp
+        return transition
+
+    # Pipeline API requirement: declare how features change (no schema change here)
+    def transform_features(self, features):
+        return features
+
+
+# Original GR00T-style collate: converts eagle_content -> eagle_* tensors
+def collate(features: list[dict[str, Any]], eagle_processor: ProcessorMixin) -> dict[str, Any]:
+    batch: dict[str, Any] = {}
+    keys = features[0].keys()
+
+    for key in keys:
+        values = [elem[key] for elem in features]
+
+        if key == "eagle_content":
+            text_list: list[str] = []
+            image_inputs: list[Any] = []
+            for v in values:
+                curr_text_list = v["text_list"]
+                curr_image_inputs = v["image_inputs"]
+                text_list += curr_text_list
+                image_inputs += curr_image_inputs
+            eagle_inputs = eagle_processor(
+                text=text_list,
+                images=image_inputs,
+                images_kwargs={"min_dynamic_tiles": 1, "max_dynamic_tiles": 1, "use_thumbnail": False},
+                return_tensors="pt",
+                padding=True,
+            )
+            for k, v in eagle_inputs.items():
+                k = "eagle_" + k
+                batch[k] = v
+        elif key in ("pixel_values", "image_grid_thw", "attention_mask", "input_ids"):
+            # Concat in existing batch dimension.
+            batch[key] = torch.cat(values)
+        else:
+            # state, state_mask, action and action_mask.
+            # Stack to form the batch dimension.
+            batch[key] = torch.from_numpy(np.stack(values))
+    return batch
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="groot_eagle_collate_v3")
+class GrootEagleCollateStep(ProcessorStep):
+    tokenizer_assets_repo: str = DEFAULT_TOKENIZER_ASSETS_REPO
+    _proc: ProcessorMixin | None = field(default=None, init=False, repr=False)
+
+    @property
+    def proc(self) -> ProcessorMixin:
+        if self._proc is None:
+            self._proc = _build_eagle_processor(self.tokenizer_assets_repo)
+        return self._proc
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        obs = transition.get(TransitionKey.OBSERVATION, {}) or {}
+        comp = transition.get(TransitionKey.COMPLEMENTARY_DATA, {}) or {}
+        contents = comp.get("eagle_content")
+        if not contents:
+            return transition
+
+        # Build features list as original API expects: one dict per batch item
+        features = [{"eagle_content": content} for content in contents]
+        batched = collate(features, self.proc)
+
+        # Inject eagle_* tensors and remove the temporary content and raw video to free memory
+        for k, v in batched.items():
+            comp[k] = v
+        comp.pop("eagle_content", None)
+        obs.pop(
+            "video", None
+        )  # The video has been fully encoded into eagle_* tensors, so we don't need the raw video anymore
+        transition[TransitionKey.OBSERVATION] = obs
+        transition[TransitionKey.COMPLEMENTARY_DATA] = comp
+        return transition
+
+    def transform_features(self, features):
+        return features
+
+
+@dataclass
+@ProcessorStepRegistry.register(name="groot_action_unpack_unnormalize_v1")
+class GrootActionUnpackUnnormalizeStep(ProcessorStep):
+    env_action_dim: int = 0
+    # Apply inverse of min-max normalization if it was used in preprocessor
+    normalize_min_max: bool = True
+    stats: dict[str, dict[str, Any]] | None = None
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        # Expect model outputs to be in TransitionKey.ACTION as (B, T, D_model)
+        action = transition.get(TransitionKey.ACTION)
+        if not isinstance(action, torch.Tensor):
+            return transition
+
+        # Select last timestep and slice to env dimension
+        if action.dim() == 3:
+            action = action[:, -1, :]
+        # Now action is (B, D_model)
+        if self.env_action_dim and action.shape[-1] >= self.env_action_dim:
+            action = action[..., : self.env_action_dim]
+
+        # Inverse min-max normalization mirroring _min_max_norm:
+        # forward: y = 2 * (x - min) / denom - 1, with y=0 when denom==0
+        # inverse: x = (y+1)/2 * denom + min, and when denom==0 -> x = min
+        if self.normalize_min_max and self.stats is not None:
+            stats_k = self.stats.get("action", {})
+            d = action.shape[-1]
+            min_v = torch.as_tensor(
+                stats_k.get("min", torch.zeros(d)), dtype=action.dtype, device=action.device
+            )
+            max_v = torch.as_tensor(
+                stats_k.get("max", torch.ones(d)), dtype=action.dtype, device=action.device
+            )
+            if min_v.numel() != d:
+                min_v = torch.nn.functional.pad(min_v.flatten()[:d], (0, max(0, d - min_v.numel())))
+                min_v = min_v.to(action.device, dtype=action.dtype)
+            if max_v.numel() != d:
+                max_v = torch.nn.functional.pad(max_v.flatten()[:d], (0, max(0, d - max_v.numel())))
+                max_v = max_v.to(action.device, dtype=action.dtype)
+            denom = max_v - min_v
+            mask = denom != 0
+            safe_denom = torch.where(mask, denom, torch.ones_like(denom))
+            inv = (action + 1.0) * 0.5 * safe_denom + min_v
+            action = torch.where(mask, inv, min_v)
+
+        transition[TransitionKey.ACTION] = action
+        return transition
+
+    def transform_features(self, features):
+        return features
+
+    def get_config(self) -> dict[str, Any]:
+        """
+        Returns a serializable dictionary of the processor's configuration.
+
+        Excludes 'stats' since they are saved separately via state_dict().
+        """
+        return {
+            "env_action_dim": self.env_action_dim,
+            "normalize_min_max": self.normalize_min_max,
+        }
+
+    def state_dict(self) -> dict[str, torch.Tensor]:
+        """
+        Returns normalization statistics as a flat state dictionary.
+
+        This enables saving stats to safetensors files, similar to normalizer_processor.
+        """
+        if not self.stats:
+            return {}
+
+        flat: dict[str, torch.Tensor] = {}
+        for key, sub in self.stats.items():
+            for stat_name, value in sub.items():
+                tensor = torch.as_tensor(value).cpu()
+                flat[f"{key}.{stat_name}"] = tensor
+        return flat
+
+    def load_state_dict(self, state: dict[str, torch.Tensor]) -> None:
+        """
+        Loads normalization statistics from a flat state dictionary.
+
+        This enables loading stats from safetensors files during from_pretrained.
+        """
+        if not state:
+            return
+
+        reconstructed: dict[str, dict[str, Any]] = {}
+        for flat_key, tensor in state.items():
+            if "." in flat_key:
+                key, stat_name = flat_key.rsplit(".", 1)
+                if key not in reconstructed:
+                    reconstructed[key] = {}
+                reconstructed[key][stat_name] = tensor
+
+        if reconstructed:
+            self.stats = reconstructed

src/lerobot/policies/groot/utils.py

@@ -0,0 +1,47 @@
+from pathlib import Path
+from shutil import copytree
+
+from huggingface_hub import hf_hub_download
+
+
+def ensure_eagle_cache_ready(vendor_dir: Path, cache_dir: Path, assets_repo: str) -> None:
+    """Populate the Eagle processor directory in cache and ensure tokenizer assets exist.
+
+    - Copies the vendored Eagle files into cache_dir (overwriting when needed).
+    - Downloads vocab.json and merges.txt into the same cache_dir if missing.
+    """
+    cache_dir = Path(cache_dir)
+    vendor_dir = Path(vendor_dir)
+
+    try:
+        # Populate/refresh cache with vendor files to ensure a complete processor directory
+        print(f"[GROOT] Copying vendor Eagle files to cache: {vendor_dir} -> {cache_dir}")
+        copytree(vendor_dir, cache_dir, dirs_exist_ok=True)
+    except Exception as exc:  # nosec: B110
+        print(f"[GROOT] Warning: Failed to copy vendor Eagle files to cache: {exc}")
+
+    required_assets = [
+        "vocab.json",
+        "merges.txt",
+        "added_tokens.json",
+        "chat_template.json",
+        "special_tokens_map.json",
+        "config.json",
+        "generation_config.json",
+        "preprocessor_config.json",
+        "processor_config.json",
+        "tokenizer_config.json",
+    ]
+
+    print(f"[GROOT] Assets repo: {assets_repo} \n Cache dir: {cache_dir}")
+
+    for fname in required_assets:
+        dst = cache_dir / fname
+        if not dst.exists():
+            print(f"[GROOT] Fetching {fname}")
+            hf_hub_download(
+                repo_id=assets_repo,
+                filename=fname,
+                repo_type="model",
+                local_dir=str(cache_dir),
+            )

src/lerobot/scripts/lerobot_eval.py

@@ -505,14 +505,17 @@ def eval_main(cfg: EvalPipelineConfig):
     )
 
     policy.eval()
+
+    # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
+    preprocessor_overrides = {
+        "device_processor": {"device": str(policy.config.device)},
+        "rename_observations_processor": {"rename_map": cfg.rename_map},
+    }
+
     preprocessor, postprocessor = make_pre_post_processors(
         policy_cfg=cfg.policy,
         pretrained_path=cfg.policy.pretrained_path,
-        # The inference device is automatically set to match the detected hardware, overriding any previous device settings from training to ensure compatibility.
-        preprocessor_overrides={
-            "device_processor": {"device": str(policy.config.device)},
-            "rename_observations_processor": {"rename_map": cfg.rename_map},
-        },
+        preprocessor_overrides=preprocessor_overrides,
     )
     with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext():
         info = eval_policy_all(

src/lerobot/utils/import_utils.py

@@ -62,6 +62,7 @@ def is_package_available(pkg_name: str, return_version: bool = False) -> tuple[b
 
 
 _transformers_available = is_package_available("transformers")
+_peft_available = is_package_available("peft")
 
 
 def make_device_from_device_class(config: ChoiceRegistry) -> Any: