Reset rollout state after robot episode end

src/lerobot/processor/relative_action_processor.py

@@ -146,6 +146,9 @@ class RelativeActionsProcessorStep(ProcessorStep):
         """Return the cached ``observation.state`` used as the reference point for relative/absolute action conversions."""
         return self._last_state
 
+    def reset(self) -> None:
+        self._last_state = None
+
     def get_config(self) -> dict[str, Any]:
         return {
             "enabled": self.enabled,

src/lerobot/rollout/robot_wrapper.py

@@ -36,6 +36,7 @@ class ThreadSafeRobot:
     def __init__(self, robot: Robot) -> None:
         self._robot = robot
         self._lock = Lock()
+        self._last_action_response: Any | None = None
 
     # -- Lock-protected I/O --------------------------------------------------
 
@@ -45,7 +46,15 @@ class ThreadSafeRobot:
 
     def send_action(self, action: dict[str, Any] | Any) -> Any:
         with self._lock:
-            return self._robot.send_action(action)
+            response = self._robot.send_action(action)
+            self._last_action_response = getattr(self._robot, "last_action_response", response)
+            return response
+
+    def pop_last_action_response(self) -> Any | None:
+        with self._lock:
+            response = self._last_action_response
+            self._last_action_response = None
+            return response
 
     # -- Read-only proxies (no lock needed) -----------------------------------
 

src/lerobot/rollout/strategies/base.py

@@ -67,6 +67,8 @@ class BaseStrategy(RolloutStrategy):
 
             action_dict = send_next_action(obs_processed, obs, ctx, interpolator)
             self._log_telemetry(obs_processed, action_dict, ctx.runtime)
+            if action_dict is not None:
+                self._reset_inference_after_robot_episode_done(ctx)
 
             dt = time.perf_counter() - loop_start
             if (sleep_t := control_interval - dt) > 0:

src/lerobot/rollout/strategies/core.py

@@ -116,6 +116,22 @@ class RolloutStrategy(abc.ABC):
             engine.resume()
         return False
 
+    def _reset_inference_after_robot_episode_done(self, ctx: RolloutContext) -> None:
+        """Reset rollout-side episode state when a robot backend reports an environment reset."""
+        response = ctx.hardware.robot_wrapper.pop_last_action_response()
+        if not isinstance(response, dict) or not response.get("done"):
+            return
+
+        logger.info(
+            "Robot reported episode done (success=%s); resetting rollout inference state",
+            response.get("success"),
+        )
+        if self._engine is not None:
+            self._engine.reset()
+        if self._interpolator is not None:
+            self._interpolator.reset()
+        self._cached_obs_processed = None
+
     def _teardown_hardware(self, hw: HardwareContext, return_to_initial_position: bool = True) -> None:
         """Stop the inference engine, optionally return robot to initial position, and disconnect hardware."""
         if self._engine is not None:

tests/policies/test_relative_actions.py

@@ -171,6 +171,23 @@ def test_full_pipeline_roundtrip(dataset, action_dim):
     torch.testing.assert_close(recovered_actions, original_actions, atol=1e-4, rtol=1e-4)
 
 
+def test_relative_actions_processor_reset_clears_cached_state():
+    relative_step = RelativeActionsProcessorStep(enabled=True)
+    transition = batch_to_transition(
+        {
+            OBS_STATE: torch.tensor([[1.0, 2.0]]),
+            ACTION: torch.tensor([[[1.5, 1.0]]]),
+        }
+    )
+
+    relative_step(transition)
+    assert relative_step.get_cached_state() is not None
+
+    relative_step.reset()
+
+    assert relative_step.get_cached_state() is None
+
+
 def test_normalized_relative_values_are_reasonable(dataset, action_dim):
     """With correct chunk stats, normalized relative actions should be in a reasonable range."""
     action_chunks, states = _build_action_chunks(dataset, CHUNK_SIZE)

tests/test_rollout.py

@@ -177,6 +177,26 @@ def test_thread_safe_robot_delegates():
     robot.disconnect()
 
 
+def test_thread_safe_robot_tracks_action_response_metadata():
+    from lerobot.rollout.robot_wrapper import ThreadSafeRobot
+
+    class MetadataRobot:
+        def get_observation(self):
+            return {}
+
+        def send_action(self, action):
+            self.last_action_response = {"action": action, "done": True, "success": False}
+            return action
+
+    robot = MetadataRobot()
+    wrapper = ThreadSafeRobot(robot)
+
+    action = {"motor_1.pos": 0.0, "motor_2.pos": 1.0, "motor_3.pos": 2.0}
+    assert wrapper.send_action(action) == action
+    assert wrapper.pop_last_action_response() == {"action": action, "done": True, "success": False}
+    assert wrapper.pop_last_action_response() is None
+
+
 def test_thread_safe_robot_properties():
     from lerobot.rollout.robot_wrapper import ThreadSafeRobot
     from tests.mocks.mock_robot import MockRobot, MockRobotConfig
@@ -194,6 +214,42 @@ def test_thread_safe_robot_properties():
     robot.disconnect()
 
 
+def test_base_strategy_resets_inference_when_robot_reports_episode_done():
+    from lerobot.rollout import BaseStrategy, BaseStrategyConfig
+
+    strategy = BaseStrategy(BaseStrategyConfig())
+    strategy._engine = MagicMock()
+    strategy._interpolator = MagicMock()
+    strategy._cached_obs_processed = {"stale": True}
+
+    ctx = MagicMock()
+    ctx.hardware.robot_wrapper.pop_last_action_response.return_value = {"done": True, "success": False}
+
+    strategy._reset_inference_after_robot_episode_done(ctx)
+
+    strategy._engine.reset.assert_called_once()
+    strategy._interpolator.reset.assert_called_once()
+    assert strategy._cached_obs_processed is None
+
+
+def test_base_strategy_ignores_action_responses_without_episode_done():
+    from lerobot.rollout import BaseStrategy, BaseStrategyConfig
+
+    strategy = BaseStrategy(BaseStrategyConfig())
+    strategy._engine = MagicMock()
+    strategy._interpolator = MagicMock()
+    strategy._cached_obs_processed = {"cached": True}
+
+    ctx = MagicMock()
+    ctx.hardware.robot_wrapper.pop_last_action_response.return_value = {"done": False}
+
+    strategy._reset_inference_after_robot_episode_done(ctx)
+
+    strategy._engine.reset.assert_not_called()
+    strategy._interpolator.reset.assert_not_called()
+    assert strategy._cached_obs_processed == {"cached": True}
+
+
 # ---------------------------------------------------------------------------
 # Strategy factory
 # ---------------------------------------------------------------------------