Fix EVO1 LIBERO eval action postprocessing

docs/source/evo1.mdx

@@ -159,14 +159,18 @@ pixel embeddings, VLM fused tokens, normalized actions, and denormalized actions
 (`max_abs_diff=0.0`).
 
 The published checkpoint expects the raw LIBERO camera feature names
-`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`. To run the converted
-checkpoint with LeRobot LIBERO evaluation for the same one-episode-per-task setting, keep those camera names
-instead of the default `image`/`image2` mapping:
+`observation.images.agentview_image` and `observation.images.robot0_eye_in_hand_image`. The official EVO1 LIBERO
+rollout protocol also replans every 14 actions and binarizes the gripper command before stepping the simulator.
+The LIBERO environment postprocessor applies the gripper binarization automatically for EVO1 policies. To run the
+converted checkpoint with LeRobot LIBERO evaluation for the same one-episode-per-task setting, keep the raw camera
+names instead of the default `image`/`image2` mapping and override `policy.n_action_steps` to 14:
 
 ```bash
 lerobot-eval \
   --policy.path=javadcc/evo1-libero-lerobot \
+  --policy.vlm_model_name=OpenGVLab/InternVL3-1B \
   --policy.device=cuda \
+  --policy.n_action_steps=14 \
   --env.type=libero \
   --env.task=libero_object \
   --env.camera_name_mapping="{agentview_image: agentview_image, robot0_eye_in_hand_image: robot0_eye_in_hand_image}" \

src/lerobot/envs/configs.py

@@ -357,6 +357,7 @@ class LiberoEnv(EnvConfig):
         }
     )
     control_mode: str = "relative"  # or "absolute"
+    binarize_gripper: bool | None = None
 
     def __post_init__(self):
         if self.obs_type == "pixels":
@@ -442,12 +443,21 @@ class LiberoEnv(EnvConfig):
         )
 
     def get_env_processors(self, policy_cfg: Any | None = None):
-        max_state_dim = getattr(policy_cfg, "max_state_dim", None) if getattr(policy_cfg, "type", None) == "evo1" else None
+        is_evo1 = getattr(policy_cfg, "type", None) == "evo1"
+        max_state_dim = getattr(policy_cfg, "max_state_dim", None) if is_evo1 else None
         action_feature = self.features.get(ACTION)
         action_dim = int(action_feature.shape[0]) if action_feature is not None else 7
+        binarize_gripper = is_evo1 if self.binarize_gripper is None else self.binarize_gripper
         return (
             PolicyProcessorPipeline(steps=[LiberoProcessorStep(max_state_dim=max_state_dim)]),
-            PolicyProcessorPipeline(steps=[LiberoActionProcessorStep(action_dim=action_dim)]),
+            PolicyProcessorPipeline(
+                steps=[
+                    LiberoActionProcessorStep(
+                        action_dim=action_dim,
+                        binarize_gripper=binarize_gripper,
+                    )
+                ]
+            ),
         )
 
 

src/lerobot/processor/env_processor.py

@@ -173,13 +173,42 @@ class LiberoActionProcessorStep(ActionProcessorStep):
     """Slices padded policy actions back to the executable LIBERO action space."""
 
     action_dim: int = 7
+    binarize_gripper: bool = False
+    gripper_index: int = 6
+    gripper_threshold: float = 0.5
+    gripper_below_threshold_value: float = 1.0
+    gripper_above_threshold_value: float = -1.0
 
     def action(self, action):
         if action.shape[-1] < self.action_dim:
             raise ValueError(
                 f"LIBERO action has {action.shape[-1]} dims, which is smaller than action_dim={self.action_dim}."
             )
-        return action[..., : self.action_dim]
+        action = action[..., : self.action_dim]
+        if not self.binarize_gripper:
+            return action
+
+        if not 0 <= self.gripper_index < self.action_dim:
+            raise ValueError(
+                f"gripper_index={self.gripper_index} must be within sliced action_dim={self.action_dim}."
+            )
+        action = action.clone()
+        below = torch.as_tensor(
+            self.gripper_below_threshold_value,
+            dtype=action.dtype,
+            device=action.device,
+        )
+        above = torch.as_tensor(
+            self.gripper_above_threshold_value,
+            dtype=action.dtype,
+            device=action.device,
+        )
+        action[..., self.gripper_index] = torch.where(
+            action[..., self.gripper_index] > self.gripper_threshold,
+            above,
+            below,
+        )
+        return action
 
     def transform_features(
         self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
@@ -190,7 +219,14 @@ class LiberoActionProcessorStep(ActionProcessorStep):
         return new_features
 
     def get_config(self) -> dict:
-        return {"action_dim": self.action_dim}
+        return {
+            "action_dim": self.action_dim,
+            "binarize_gripper": self.binarize_gripper,
+            "gripper_index": self.gripper_index,
+            "gripper_threshold": self.gripper_threshold,
+            "gripper_below_threshold_value": self.gripper_below_threshold_value,
+            "gripper_above_threshold_value": self.gripper_above_threshold_value,
+        }
 
 
 @dataclass

tests/envs/test_dispatch.py

@@ -99,12 +99,18 @@ def test_libero_evo1_processors_use_padded_state_and_env_action_dim():
     assert pre.steps[0].max_state_dim == 24
     assert isinstance(post.steps[0], LiberoActionProcessorStep)
     assert post.steps[0].action_dim == cfg.features["action"].shape[0] == 7
+    assert post.steps[0].binarize_gripper is True
 
     class _OtherConfig:
         type = "other"
 
-    pre_other, _ = make_env_pre_post_processors(cfg, policy_cfg=_OtherConfig())
+    pre_other, post_other = make_env_pre_post_processors(cfg, policy_cfg=_OtherConfig())
     assert pre_other.steps[0].max_state_dim is None
+    assert post_other.steps[0].binarize_gripper is False
+
+    cfg.binarize_gripper = False
+    _, post_disabled = make_env_pre_post_processors(cfg, policy_cfg=_Evo1Config())
+    assert post_disabled.steps[0].binarize_gripper is False
 
 
 def test_libero_processor_pads_state_to_max_dim():
@@ -138,6 +144,24 @@ def test_libero_action_processor_slices_padded_action():
         step.action(torch.zeros(2, 6))
 
 
+def test_libero_action_processor_can_binarize_gripper():
+    step = LiberoActionProcessorStep(action_dim=7, binarize_gripper=True)
+    action = torch.tensor(
+        [
+            [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 0.5, 7.0],
+            [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 0.6, 7.0],
+        ],
+        dtype=torch.float32,
+    )
+
+    processed = step.action(action)
+
+    assert processed.shape == (2, 7)
+    assert torch.equal(processed[:, :6], action[:, :6])
+    assert torch.equal(processed[:, 6], torch.tensor([1.0, -1.0]))
+    assert torch.equal(action[:, 6], torch.tensor([0.5, 0.6]))
+
+
 def test_base_create_envs():
     """Base class create_envs() should build a single-task VectorEnv via gym.make()."""
     gym_id = "_dispatch_test/CartPole-v99"