Fix gym-hil integration with the new LeRobot pipeline. (#2482)

* Add GymHILAdapterProcessorStep for gym-hil environment integration

* Fix action features in control loop for None teleop device with gym-hil

* Finalize dataset before pushing to hub for visualization on the hub

* Fix neutral action for gripper

* fix pre-commit

src/lerobot/processor/__init__.py

@@ -44,6 +44,7 @@ from .hil_processor import (
     AddTeleopActionAsComplimentaryDataStep,
     AddTeleopEventsAsInfoStep,
     GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     RewardClassifierProcessorStep,
@@ -87,6 +88,7 @@ __all__ = [
     "DoneProcessorStep",
     "EnvAction",
     "EnvTransition",
+    "GymHILAdapterProcessorStep",
     "GripperPenaltyProcessorStep",
     "hotswap_stats",
     "IdentityProcessorStep",

src/lerobot/processor/gym_action_processor.py

@@ -20,6 +20,7 @@ from lerobot.configs.types import PipelineFeatureType, PolicyFeature
 
 from .converters import to_tensor
 from .core import EnvAction, EnvTransition, PolicyAction
+from .hil_processor import TELEOP_ACTION_KEY
 from .pipeline import ActionProcessorStep, ProcessorStep, ProcessorStepRegistry
 
 
@@ -89,6 +90,13 @@ class Numpy2TorchActionProcessorStep(ProcessorStep):
             torch_action = to_tensor(action, dtype=None)  # Preserve original dtype
             new_transition[TransitionKey.ACTION] = torch_action
 
+        complementary_data = new_transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+        if TELEOP_ACTION_KEY in complementary_data:
+            teleop_action = complementary_data[TELEOP_ACTION_KEY]
+            if isinstance(teleop_action, EnvAction):
+                complementary_data[TELEOP_ACTION_KEY] = to_tensor(teleop_action)
+            new_transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
         return new_transition
 
     def transform_features(

src/lerobot/processor/hil_processor.py

@@ -312,6 +312,37 @@ class TimeLimitProcessorStep(TruncatedProcessorStep):
         return features
 
 
+@ProcessorStepRegistry.register("gym_hil_adapter_processor")
+class GymHILAdapterProcessorStep(ProcessorStep):
+    """
+    Adapts the output of the `gym-hil` environment to the format expected by `lerobot` processors.
+
+    This step normalizes the `transition` object by:
+    1. Copying `teleop_action` from `info` to `complementary_data`.
+    2. Copying `is_intervention` from `info` (using the string key) to `info` (using the enum key).
+    """
+
+    def __call__(self, transition: EnvTransition) -> EnvTransition:
+        info = transition.get(TransitionKey.INFO, {})
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA, {})
+
+        if TELEOP_ACTION_KEY in info:
+            complementary_data[TELEOP_ACTION_KEY] = info[TELEOP_ACTION_KEY]
+
+        if "is_intervention" in info:
+            info[TeleopEvents.IS_INTERVENTION] = info["is_intervention"]
+
+        transition[TransitionKey.INFO] = info
+        transition[TransitionKey.COMPLEMENTARY_DATA] = complementary_data
+
+        return transition
+
+    def transform_features(
+        self, features: dict[PipelineFeatureType, dict[str, PolicyFeature]]
+    ) -> dict[PipelineFeatureType, dict[str, PolicyFeature]]:
+        return features
+
+
 @dataclass
 @ProcessorStepRegistry.register("gripper_penalty_processor")
 class GripperPenaltyProcessorStep(ProcessorStep):

src/lerobot/rl/gym_manipulator.py

@@ -36,6 +36,7 @@ from lerobot.processor import (
     DeviceProcessorStep,
     EnvTransition,
     GripperPenaltyProcessorStep,
+    GymHILAdapterProcessorStep,
     ImageCropResizeProcessorStep,
     InterventionActionProcessorStep,
     MapDeltaActionToRobotActionStep,
@@ -379,6 +380,7 @@ def make_processors(
         ]
 
         env_pipeline_steps = [
+            GymHILAdapterProcessorStep(),
             Numpy2TorchActionProcessorStep(),
             VanillaObservationProcessorStep(),
             AddBatchDimensionProcessorStep(),
@@ -608,7 +610,14 @@ def control_loop(
 
     dataset = None
     if cfg.mode == "record":
-        action_features = teleop_device.action_features
+        if teleop_device:
+            action_features = teleop_device.action_features
+        else:
+            action_features = {
+                "dtype": "float32",
+                "shape": (4,),
+                "names": ["delta_x", "delta_y", "delta_z", "gripper"],
+            }
         features = {
             ACTION: action_features,
             REWARD: {"dtype": "float32", "shape": (1,), "names": None},
@@ -656,7 +665,7 @@ def control_loop(
         # Create a neutral action (no movement)
         neutral_action = torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
         if use_gripper:
-            neutral_action = torch.cat([neutral_action, torch.tensor([1.0])])  # Gripper stay
+            neutral_action = torch.cat([neutral_action, torch.tensor([0.0])])  # Gripper stay
 
         # Use the new step function
         transition = step_env_and_process_transition(
@@ -725,6 +734,8 @@ def control_loop(
         precise_sleep(max(dt - (time.perf_counter() - step_start_time), 0.0))
 
     if dataset is not None and cfg.dataset.push_to_hub:
+        logging.info("Finalizing dataset before pushing to hub")
+        dataset.finalize()
         logging.info("Pushing dataset to hub")
         dataset.push_to_hub()