nit

src/lerobot/policies/rlearn/configuration_rlearn.py

@@ -59,7 +59,7 @@ class RLearNConfig(PreTrainedConfig):
     use_tanh_head: bool = False  # when True, bound outputs in [-1, 1]
 
     # Training
-    learning_rate: float = 5e-5  # Reduced for stability
+    learning_rate: float = 1e-4
     weight_decay: float = 0.01
     loss_type: str = "composite"  # Always use composite loss with spatial awareness
     ranking_margin: float = 0.1

src/lerobot/policies/rlearn/modeling_rlearn.py

@@ -384,7 +384,10 @@ class RLearNPolicy(PreTrainedPolicy):
             target = torch.cat([padding, target], dim=1)  # Prepend padding
 
             import logging
-            logging.debug(f"Padded targets from {target.shape[1] - padding_needed} to {self.config.max_seq_len}")
+
+            logging.debug(
+                f"Padded targets from {target.shape[1] - padding_needed} to {self.config.max_seq_len}"
+            )
 
         # Now safely index with idx
         target = target[:, idx]
@@ -682,6 +685,7 @@ def extract_visual_sequence(batch: dict[str, Tensor], target_seq_len: int = None
             frames = torch.cat([padding, frames], dim=1)  # Prepend padding
 
             import logging
+
             logging.debug(f"Padded sequence from {T} to {target_seq_len} frames by repeating first frame")
 
     return frames

src/lerobot/policies/rlearn/rlearn_plan.md

@@ -132,7 +132,7 @@ Default weights: $\lambda_{\text{prog}}=1.0$, $\lambda_{\text{spatial-nce}}=0.5$
 - Try DINO v3 as encoder Base 86 M: https://huggingface.co/facebook/dinov3-vitb16-pretrain-lvd1689m with HuggingFaceTB/SmolLM2-135M-Instruct ? []
 - Add more artificial text to dataset generated by vlm (google gemini) []
   - See google gemini vlm caption [] https://gemini.google.com/app/7e332ffaf32580f2
-  - Multiple captions per video, creat method to generate as much data as possible etc [] https://arxiv.org/abs/2508.13446
+  - Multiple captions per video, creat method to generate as much data as possible etc [] https://arxiv.org/abs/2508.13446, https://arxiv.org/pdf/2412.04453
 - How can we improve spatial aware learning? co generating captions for each frame with language decoder?
 - Add droid []
 - Extend evaluation []