Guard GR00T relative action stepwise decode

src/lerobot/policies/groot/modeling_groot.py

@@ -468,6 +468,14 @@ class GrootPolicy(PreTrainedPolicy):
     @torch.no_grad()
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select single action from action queue."""
+        if getattr(self.config, "use_relative_actions", False):
+            raise NotImplementedError(
+                "GrootPolicy.select_action does not support relative-action policies because cached "
+                "relative chunk actions can be decoded against newer observation states. Use "
+                "predict_action_chunk and postprocess the full chunk before queuing actions, or use "
+                "the RTC/chunked rollout inference path."
+            )
+
         self.eval()
 
         if len(self._action_queue) == 0:

src/lerobot/policies/groot/processor_groot.py

@@ -2158,6 +2158,12 @@ class GrootN17ActionDecodeStep(ProcessorStep):
             return transition
 
         action_np = action.detach().cpu().float().numpy()
+        if self.use_relative_action and action_np.ndim != 3:
+            raise NotImplementedError(
+                "GrootN17ActionDecodeStep cannot decode native relative actions one step at a time. "
+                "Decode the full action chunk returned by predict_action_chunk while the matching "
+                "GrootN17PackInputsStep state is still cached, then queue the decoded absolute actions."
+            )
         # The sync action queue postprocesses popped actions as (B, D); decode
         # them as single-step (B, 1, D) chunks and squeeze the horizon back at
         # the end so both ranks share the chunk decode logic below.