make it work

eval.sh

@@ -1,5 +1,5 @@
 lerobot-eval \
-  --policy.path="/raid/jade/models/xvla-lib" \
+  --policy.path="/raid/jade/models/xvla-libero-new_migrated2" \
   --env.type=libero \
   --env.task=libero_spatial \
   --env.control_mode=absolute \

src/lerobot/envs/libero.py

@@ -244,7 +244,7 @@ class LiberoEnv(gym.Env):
         eef_quat = raw_obs.get("robot0_eef_quat")
 
         # rotation matrix from controller
-        eef_mat = self._env.robots[0].controller.ee_ori_mat if eef_pos is not None else None
+        eef_mat = self._env.robots[0].controller.ee_ori_mat
         gripper_qpos = raw_obs.get("robot0_gripper_qpos")
         gripper_qvel = raw_obs.get("robot0_gripper_qvel")
         joint_pos = raw_obs.get("robot0_joint_pos")

src/lerobot/policies/xvla/__init__.py

@@ -0,0 +1,6 @@
+from lerobot.policies.xvla.processor_xvla import (
+    XVLAAddDomainIdProcessorStep,
+    XVLAImageScaleProcessorStep,
+    XVLARotation6DToAxisAngleProcessorStep,
+    make_xvla_pre_post_processors,
+)

src/lerobot/policies/xvla/compare.py

@@ -0,0 +1,43 @@
+import numpy as np
+import os
+
+# --- CHANGE THESE TWO PATHS ---
+dir_a = "/home/jade_choghari/robot/lerobot"
+dir_b = "/home/jade_choghari/robot/robot2/lerobot"
+
+# keys = ["input_ids", "image_input", "image_mask", "domain_id", "proprio"]
+keys = ["domain_id"]
+
+def load_np(path, key):
+    file_path = os.path.join(path, f"{key}.npy")
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"Missing file: {file_path}")
+    return np.load(file_path)
+
+for k in keys:
+    print(f"\n===== {k} =====")
+    a = load_np(dir_a, k)
+    b = load_np(dir_b, k)
+    print(a)
+    print(b)
+    print("Shapes:", a.shape, b.shape)
+    same_shape = a.shape == b.shape
+    print("Same shape:", same_shape)
+
+    if not same_shape:
+        continue
+
+    # Allclose for numeric, array_equal for boolean
+    if a.dtype == bool or b.dtype == bool:
+        equal = np.array_equal(a, b)
+        print("Equal:", equal)
+
+        # For boolean arrays, diff = xor
+        diff = np.sum(a ^ b)
+        print("Num differing elements:", diff)
+    else:
+        close = np.allclose(a, b, atol=1e-6, rtol=1e-6)
+        print("Allclose:", close)
+
+        diff = np.max(np.abs(a - b))
+        print("Max difference:", diff)

src/lerobot/policies/xvla/processor_xvla.py

@@ -62,6 +62,7 @@ def make_xvla_pre_post_processors(
             padding_side=config.tokenizer_padding_side,
         ),
         DeviceProcessorStep(device=config.device),
+        XVLAAddDomainIdProcessorStep(),
         NormalizerProcessorStep(
             features=features, norm_map=config.normalization_mapping, stats=dataset_stats
         ),
@@ -276,7 +277,7 @@ class XVLAAddDomainIdProcessorStep(ProcessorStep):
         device: Device to place the domain_id tensor on (default: "cuda")
     """
 
-    domain_id: int = 3
+    domain_id: int = 0
     device: str = "cuda"
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
@@ -293,7 +294,7 @@ class XVLAAddDomainIdProcessorStep(ProcessorStep):
                 if isinstance(v, torch.Tensor):
                     batch_size = v.shape[0]
                     break
-
+    
         # Add domain_id tensor
         comp["domain_id"] = torch.tensor([int(self.domain_id)] * batch_size, dtype=torch.long).to(self.device)
 
@@ -387,7 +388,7 @@ def make_xvla_libero_pre_post_processors(
     """
     pre_processor_steps: list[ProcessorStep] = []
     post_processor_steps: list[ProcessorStep] = []
-    pre_processor_steps.extend([LiberoProcessorStep(), XVLAImageScaleProcessorStep(), XVLAAddDomainIdProcessorStep()])
+    pre_processor_steps.extend([LiberoProcessorStep(), XVLAAddDomainIdProcessorStep()])
     post_processor_steps.extend([XVLARotation6DToAxisAngleProcessorStep()])
     return (
         PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](

src/lerobot/processor/normalize_processor.py

@@ -338,7 +338,7 @@ class _NormalizationMixin:
                 return (tensor * std + mean) * 255.0
 
             # Normalize
-            return (tensor / 255.0 - mean) / std
+            return (tensor - mean) / std
 
         stats = self._tensor_stats[key]
         if norm_mode == NormalizationMode.MEAN_STD:

src/lerobot/scripts/lerobot_eval.py

@@ -167,10 +167,8 @@ def rollout(
         # Infer "task" from attributes of environments.
         # TODO: works with SyncVectorEnv but not AsyncVectorEnv
         observation = add_envs_task(env, observation)
-
         # Apply environment-specific preprocessing (e.g., LiberoProcessorStep for LIBERO)
         observation = env_preprocessor(observation)
-
         observation = preprocessor(observation)
         with torch.inference_mode():
             action = policy.select_action(observation)
@@ -178,6 +176,7 @@ def rollout(
         action_transition = {"action": action}
         action_transition = env_postprocessor(action_transition)
         action = action_transition["action"]
+
         # Convert to CPU / numpy.
         action_numpy: np.ndarray = action.to("cpu").numpy()
         assert action_numpy.ndim == 2, "Action dimensions should be (batch, action_dim)"