From f84c20d403cbfd3f1f96f4a6e7b2d43cb853a716 Mon Sep 17 00:00:00 2001 From: Pepijn Date: Mon, 1 Sep 2025 11:59:20 +0200 Subject: [PATCH] huberman loss --- src/lerobot/policies/rlearn/modeling_rlearn.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/lerobot/policies/rlearn/modeling_rlearn.py b/src/lerobot/policies/rlearn/modeling_rlearn.py index 1940c0bc1..2f86b88ed 100644 --- a/src/lerobot/policies/rlearn/modeling_rlearn.py +++ b/src/lerobot/policies/rlearn/modeling_rlearn.py @@ -522,8 +522,8 @@ class RLearNPolicy(PreTrainedPolicy): if mismatch_tensor.any(): eps = self.config.logit_eps zeros_target_logits = torch.logit(torch.full_like(mismatch_raw_logits, eps)) - mismatch_loss_per_sample = F.mse_loss( - mismatch_raw_logits, zeros_target_logits, reduction='none' + mismatch_loss_per_sample = F.smooth_l1_loss( + mismatch_raw_logits, zeros_target_logits, beta=0.5, reduction='none' ).mean(dim=1) L_mismatch = mismatch_loss_per_sample[mismatch_tensor].mean()