Commit Graph

  • b8ad81bf39 feat(rewards): add ROBOMETER reward model (#3627) main Khalil Meftah 2026-05-29 21:45:39 +02:00
  • ac9128b319 fixing simlink feat/vla-jepa Maximellerbach 2026-05-29 15:18:24 +02:00
  • 61d82a5773 adding configuration gripper index and threshold Maximellerbach 2026-05-29 14:11:51 +02:00
  • d8ed30d58c removing swish in favor of silu Maximellerbach 2026-05-29 13:10:21 +02:00
  • 79f6756505 cleanup Maximellerbach 2026-05-29 12:53:40 +02:00
  • 7a98c56ce4 removing useless pre-processor Maximellerbach 2026-05-29 12:28:49 +02:00
  • 9961eb6918 adding .mdx docs and shortening polivy_vla_jepa_README.md Maximellerbach 2026-05-29 11:52:42 +02:00
  • e1852db71a adding licences Maximellerbach 2026-05-29 11:51:58 +02:00
  • c2b29c8ae0 removing conversion script Maximellerbach 2026-05-28 11:47:00 +02:00
  • 58eac863aa fixing misconception about multiview / singleview handling Maximellerbach 2026-05-27 11:16:37 +02:00
  • 952e5146dc smol fix to avoid having default CPU device when training Maximellerbach 2026-05-26 15:03:29 +02:00
  • 37fda2a6fc adding instructions for different embodiement + fixing some tests Maximellerbach 2026-05-26 11:51:54 +02:00
  • df7d5132d1 fix qwen norm layer output libero eval is now as expected Maximellerbach 2026-05-22 17:36:43 +02:00
  • 8efa5cabe9 trying to close success rate gap Maxime Ellerbach 2026-05-22 07:59:47 +00:00
  • 7e23859c55 fixing training and exal examples Maximellerbach 2026-05-21 16:35:50 +02:00
  • a24f669deb adding guard for diffusers Maximellerbach 2026-05-21 15:41:26 +02:00
  • b7727b8a6c adressing dtype zeros issue Maximellerbach 2026-05-21 15:40:17 +02:00
  • 7db4414e6b fixing doc defaults args Maxime Ellerbach 2026-05-21 15:37:37 +02:00
  • 7da594fda8 pre-commit cleanup Maximellerbach 2026-05-21 13:16:57 +02:00
  • 01ce5d7af1 refactoring into using pre and post processor Maxime Ellerbach 2026-05-21 11:15:52 +00:00
  • 83ef59e020 lots of changes to make existing weights work, need to massively refactor the pre and post processing Maxime Ellerbach 2026-05-20 15:24:59 +00:00
  • 997b713f14 removing missleading future_action_window_size to just use chunk_size Maximellerbach 2026-05-18 18:14:13 +02:00
  • 1e3d25f10e allow different state dim and action dim Maximellerbach 2026-05-18 14:32:32 +02:00
  • 8c0efb8295 trying out to re-init the action head to avoid pretraining dimension mismatch Maximellerbach 2026-05-18 13:54:07 +02:00
  • 9ef0fd5433 make default params more aligned with paper and pretrained models - adding possibility of freezing qwen backbone and world model - added tests for weight loading Maximellerbach 2026-05-18 12:55:55 +02:00
  • 26d2ac48a8 add one-shot script to convert ginwind/VLA-JEPA checkpoints to safetensors (will remove once migrated) Maximellerbach 2026-05-15 15:59:43 +02:00
  • 0f29cd3167 add VLA-JEPA documentation Maximellerbach 2026-05-15 14:42:40 +02:00
  • 64c9570547 update VLA-JEPA tests for arch changes and action_is_pad Maximellerbach 2026-05-15 14:34:41 +02:00
  • 8b03d25fef propagate action_is_pad masking through VLA-JEPA policy pipeline Maximellerbach 2026-05-15 14:34:21 +02:00
  • 5e37d97631 align VLA-JEPA architecture with original checkpoint Maximellerbach 2026-05-15 14:34:10 +02:00
  • a71e0d34ad adding more tests to ensure good coverage Maximellerbach 2026-05-13 15:55:04 +02:00
  • d00b3e993a some more fixes to be closer to the original implem Maximellerbach 2026-05-13 15:51:55 +02:00
  • 596c72bfc6 adjusting obs steps, tublets size to match original implementation Maxime Ellerbach 2026-05-13 12:38:42 +00:00
  • ea535ad98d fixing wm_loss not propagating Maxime Ellerbach 2026-05-13 12:38:28 +00:00
  • 7368a0085a fix warnings with qwen processor kwargs Maxime Ellerbach 2026-05-13 08:34:38 +00:00
  • 7dba4f19a9 fixing action and state dim Maxime Ellerbach 2026-05-13 08:33:27 +00:00
  • 90d398ea59 adding guards to avoid needing transformers and diffusers for type checking and basic tests Maximellerbach 2026-05-11 19:05:26 +02:00
  • 16a4643000 updating uv lock Maximellerbach 2026-05-11 17:03:08 +02:00
  • 6fa78ca8b4 adding deps to pyproject.toml Maximellerbach 2026-05-11 16:59:14 +02:00
  • ec4bf4e47f linting Maximellerbach 2026-05-11 16:49:18 +02:00
  • da56489174 (feat)policies: add VLA-JEPA ginwind 2026-05-11 06:54:15 +00:00
  • addb354296 support vla_jepa ginwind 2026-05-01 11:39:06 +00:00
  • 848ed3240e feat(policies): add VLA-JEPA ginwind 2026-04-14 04:03:34 +00:00
  • f7bb1795e7 feat(policies): add VLA-JEPA ginwind 2026-04-02 07:00:44 +00:00
  • 0355902fba first commit ginwind 2026-04-02 06:28:34 +00:00
  • badbb69fb7 fix(rewards/topreward): Remove TOPReward README symlink, use docs page only fix/topreward-docs-symlink Khalil Meftah 2026-05-27 18:50:14 +01:00
  • 23036baf22 fix(rewards/topreward): dd README symlink for TOPReward docs Khalil Meftah 2026-05-27 18:33:00 +01:00
  • 24017e960c Add MolmoAct2 policy (#3604) Haoquan Fang 2026-05-27 09:58:37 -07:00
  • e86f5af5bf feat(rewards): add TOPReward reward model (#3629) Khalil Meftah 2026-05-27 14:24:31 +02:00
  • 1f1541243a pi052: make `lerobot-eval` work on saved checkpoints feat/smolvla-on-steerable pepijn 2026-05-27 09:14:34 +00:00
  • 6b647fee2c chore(dependencies): update uv.lock auto/update-uv-lock github-actions[bot] 2026-05-27 05:00:12 +00:00
  • 5c98e80430 fix(gr00t): fix Eagle25VL model and processor crash in transformers>=5.4.0, <5.6.0 (#3652) Haoming Song 2026-05-26 20:04:22 +08:00
  • f65f3f7a4a Fix policy.path in YAML configs (PR #3145 followup) (#3597) Reece O'Mahoney 2026-05-26 13:01:19 +01:00
  • d70c810416 pi052: drop `use_hf_kernels` flag — always patch Liger kernels pepijn 2026-05-26 11:47:49 +00:00
  • 4c3ddb1ff5 pi052: wire Liger fused linear CE + DDP-safe FAST tokenizer fit pepijn 2026-05-26 11:18:16 +00:00
  • a24d10f5bb add quick AI draft for quickstart docs/complete-docs-redesign Nikodem Bartnik 2026-05-26 13:10:24 +02:00
  • bec7d668a6 chore: enable Dependabot weekly GitHub Actions bumps chore/add-dependabot-github-actions hf-dependantbot-rollout[bot] 2026-05-26 10:32:49 +00:00
  • 32279544ea add new docs chapters structure Nikodem Bartnik 2026-05-26 12:01:33 +02:00
  • 8615f3f613 annotate(vqa): tighten bbox + keypoint quality bar pepijn 2026-05-26 08:31:37 +00:00
  • 1e7c0d6aa1 annotate(plan): force composite-action subtasks; ban ultra-fine splits feat/language-annotation-pipeline pepijn 2026-05-26 05:14:30 +00:00
  • 2686450d68 annotate(plan): force composite-action subtasks; tune run_hf_job for robocasa_smoke pepijn 2026-05-26 05:14:23 +00:00
  • 920c6ef5a2 docs(annotate): disable phase-0 vocabulary discovery by default in run_hf_job pepijn 2026-05-26 04:42:10 +00:00
  • 4913356564 pi052: SDPA attention port + selective AC + bench harness pepijn 2026-05-25 21:59:20 +00:00
  • 673cc6b0fe pi052: opt-in Liger fused kernels (rope + geglu + layer_norm) pepijn 2026-05-25 20:44:02 +00:00
  • 2ed6519a93 ema: enable by default (matches openpi JAX behavior) Pepijn 2026-05-25 21:58:46 +02:00
  • 72ea531017 train: switch EMA from custom ModelEMA to ema-pytorch Pepijn 2026-05-25 21:51:23 +02:00
  • 56a934ec55 train: EMA of policy parameters (opt-in via --ema.enable=true) Pepijn 2026-05-25 21:27:14 +02:00
  • 738e317caa pi052: PaLM-style z-loss on text CE (default weight 1e-4) Pepijn 2026-05-25 21:08:56 +02:00
  • 8ba3b187a1 pi052: bump lm_head_lr_scale default to 5.0 (keep base LR at 2.5e-5) Pepijn 2026-05-25 20:57:43 +02:00
  • 057c794ffe wandb: flip training-example logging defaults to on (every 5000 steps) Pepijn 2026-05-25 18:00:04 +02:00
  • b1e83f556c train: periodic wandb log of training examples (images + text + actions) Pepijn 2026-05-25 16:57:15 +02:00
  • da3e87ee86 Merge branch 'feat/smolvla-on-steerable' of https://github.com/huggingface/lerobot into feat/smolvla-on-steerable Pepijn 2026-05-25 16:56:50 +02:00
  • 1e9a6d044d Merge remote-tracking branch 'origin/feat/language-annotation-pipeline' into feat/smolvla-on-steerable Pepijn 2026-05-25 16:56:22 +02:00
  • 3fdfcb912a examples(port_datasets): generalize RoboCasa builder + add smoke script pepijn 2026-05-25 14:51:09 +00:00
  • 3b5b94dbd6 optmize topreward input processing (#3660) Haoming Song 2026-05-25 22:07:45 +08:00
  • c37b1fc7d0 Merge origin/feat/language-annotation-pipeline (8 fix(annotate) commits + vocabulary phase) Pepijn 2026-05-25 15:47:25 +02:00
  • 9020635b14 Merge branch 'main' into feat/language-annotation-pipeline Pepijn 2026-05-25 15:46:32 +02:00
  • 83d0c390da pi052: drop debug scaffolding left over from training/inference bug hunts Pepijn 2026-05-25 15:07:43 +02:00
  • 1ff10b935c Merge branch 'feat/language-annotation-pipeline' into feat/smolvla-on-steerable Pepijn 2026-05-25 14:47:09 +02:00
  • 67bdf4690e examples(port_datasets): rewrite RoboCasa composite_seen builder Pepijn 2026-05-25 14:27:42 +02:00
  • 8085feab6e pi052(runtime): factor out shared observation-prep boilerplate Pepijn 2026-05-25 14:25:08 +02:00
  • a088c10c80 examples(port_datasets): SLURM+datatrove RoboCasa composite_seen build Pepijn 2026-05-25 14:10:05 +02:00
  • 9c3d5ab7ce scripts: build_robocasa_composite_seen — aggregate 16 target tasks Pepijn 2026-05-25 14:01:28 +02:00
  • e84f97a8c1 smolvla2(runtime): interactive task picker + drop action diagnostic Pepijn 2026-05-25 12:59:08 +02:00
  • 6d2b8c80ab smolvla2(runtime): wire MemoryUpdateFwd into the inference pipeline Pepijn 2026-05-25 12:52:44 +02:00
  • 793c7c4ddd feat(runtime): --subtask_chunks_per_gen throttles HL gen vs action chunks Pepijn 2026-05-25 12:34:59 +02:00
  • db927ab40b feat(runtime): action chunk diagnostic — log normalized + unnormalized values Pepijn 2026-05-25 12:10:52 +02:00
  • 471b2b1b1d fix(annotate): bump same-frame subtasks onto distinct frames pepijn 2026-05-23 19:31:44 +00:00
  • a15e16c072 fix(annotate): replace fuzzy subtask snapping with strict match + one-shot retry pepijn 2026-05-23 09:57:27 +00:00
  • 616663cd9f fix(rewards/topreward): fix pyproject extra typo and simplify processor (#3653) Cole 2026-05-22 17:27:09 -05:00
  • 3ab08a5318 fix(imports): fixing av import in test_depth.py feat/depth-integration Caroline Pascal 2026-05-22 15:13:15 +02:00
  • 336af85c09 fix(annotate): never leave an episode with zero canonical subtasks pepijn 2026-05-22 12:44:03 +00:00
  • 54221ceea2 feat(annotate): let the VLM decide vocabulary size pepijn 2026-05-22 11:46:31 +00:00
  • 369ab17110 fix(annotate): update run_hf_job CLI args for renamed namespaces + phase 0 pepijn 2026-05-22 11:43:06 +00:00
  • 86a7edc590 feat(annotate): phase 0 — derive canonical vocabulary from sample episodes pepijn 2026-05-22 11:40:05 +00:00
  • 5e53e6bd2f tests(typos): fixing typos in tests CarolinePascal 2026-05-22 13:09:56 +02:00
  • 77a16db529 fix(smolvla2): make HighLevelSubtaskFwd actually fire at low hz + quiet startup log pepijn 2026-05-22 11:04:12 +00:00
  • 8194897994 fix(deps): cap placo below 0.9.16 and harden kinematics import (#3647) Pepijn 2026-05-22 12:03:07 +02:00
  • ca1b951e7b feat(pi05): expose lm_head_lr_scale for stronger text-CE gradient pepijn 2026-05-22 09:56:46 +00:00
  • 9d30d91021 fix(pi052,smolvla2): unblock text generation when LM head drifted to <loc> pepijn 2026-05-22 09:50:14 +00:00