fix

src/lerobot/policies/rlearn/modeling_rlearn.py

@@ -314,13 +314,15 @@ class RLearNPolicy(PreTrainedPolicy):
         if flat.max() > 1.0:
             flat = flat / 255.0
             
-        # DINOv3 expects images in [0, 1] range, RGB format
+        # Prepare inputs for processor: feed uint8 to avoid double-rescale; let processor rescale/normalize
-        # Convert tensor to list of PIL-like arrays for processor
+        flat_clamped = flat.clamp(0.0, 1.0)
-        flat_numpy = flat.permute(0, 2, 3, 1).cpu().numpy()  # (BT, H, W, C)
+        flat_uint8 = (flat_clamped * 255.0).round().to(torch.uint8)  # (BT, C, H, W)
+        flat_numpy = flat_uint8.permute(0, 2, 3, 1).cpu().numpy()  # (BT, H, W, C) uint8
         images_list = [flat_numpy[i] for i in range(B * T)]
             
         # Process through DINOv3 processor and model
-        inputs = self.vision_processor(images=images_list, return_tensors="pt")
+        # Disable center-crop to preserve motion near borders; allow default rescale/normalize
+        inputs = self.vision_processor(images=images_list, return_tensors="pt", do_center_crop=False)
         inputs = {k: v.to(device) for k, v in inputs.items()}
         
         # Process in batch through DINOv3 model
@@ -412,6 +414,27 @@ class RLearNPolicy(PreTrainedPolicy):
                 feature_std = vision_features.std().item()
                 print(f"Feature stats: mean={feature_mean:.4f}, std={feature_std:.4f}")
 
+                # Extra DIAGNOSTIC: CLS vs patch mean/max deltas for one sample, two far-apart frames
+                try:
+                    if 'last_hidden_state' in vision_outputs.__dict__ and T >= 2:
+                        # Recover CLS tokens
+                        cls_flat = tokens[:, 0, :]  # (BT, D)
+                        cls = rearrange(cls_flat, '(b t) d -> b t d', b=B, t=T)
+                        b0 = 0
+                        f0, f1 = 0, T - 1
+                        # L2 between CLS at two frames
+                        cls_l2 = (cls[b0, f1] - cls[b0, f0]).pow(2).sum().sqrt().item()
+                        # Patch mean L2
+                        pm_f0 = patch_features[b0, f0].mean(dim=0)
+                        pm_f1 = patch_features[b0, f1].mean(dim=0)
+                        pm_l2 = (pm_f1 - pm_f0).pow(2).sum().sqrt().item()
+                        # Max over patches L2
+                        per_patch_l2 = (patch_features[b0, f1] - patch_features[b0, f0]).pow(2).sum(dim=1).sqrt()
+                        max_p_l2 = per_patch_l2.max().item()
+                        print(f"CLS ΔL2: {cls_l2:.6f} | mean(patches) ΔL2: {pm_l2:.6f} | max(patch) ΔL2: {max_p_l2:.6f}")
+                except Exception as _:
+                    pass
+                
                 # Check temporal variance for each sample
                 for b_idx in range(min(B, 2)):  # Debug first 2 samples
                     sample_features = vision_features[b_idx]  # (T, D)