Remove Arena's Gr00t Docker; use Isaac-Gr00t clients

isaaclab_arena/evaluation/policy_runner.py

@@ -204,18 +204,10 @@ def main():
             # Each rank prints its own metrics as it can be different due to random seed
             print(f"[Rank {local_rank}/{world_size}] Metrics: {metrics}")
 
-        # NOTE(huikang, 2025-12-30)Explicitly clean up the remote policy client / server.
-        # Do NOT rely on a __del__ destructor in policy for this, since destructors are
-        # triggered implicitly and their execution time (or even whether they run)
-        # is not guaranteed, which makes resource cleanup unreliable.
-        if policy.is_remote:
-            policy.shutdown_remote(kill_server=args_cli.remote_kill_on_exit)
-
         # Close the environment.
         env.close()
 
 
 if __name__ == "__main__":
-    # TODO(xinjie.yao, 2026.03.31): Remove it after policy sever-client is implemented properly in v0.3.
     ensure_groot_deps_in_path()
     main()

isaaclab_arena/policy/__init__.py

@@ -3,8 +3,8 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-from .action_chunking import ActionChunkingState
-from .action_chunking_client import *
+from .action_scheduler import ActionScheduler
+from .action_chunking import ActionChunkScheduler, ActionChunkingState
 from .replay_action_policy import *
 from .rsl_rl_action_policy import *
 from .zero_action_policy import *

isaaclab_arena/policy/action_chunking.py

@@ -3,19 +3,21 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-"""Shared action chunking state and logic for local and remote policies."""
+"""ActionChunkScheduler: buffer a model chunk and step through it sequentially."""
 
 from __future__ import annotations
 
 import torch
 from collections.abc import Callable
 
+from isaaclab_arena.policy.action_scheduler import ActionScheduler
 
-class ActionChunkingState:
-    """Holds chunk buffer, per-env index, and refill flag; provides get_action(fetch_chunk_fn).
 
-    Used by both Gr00tClosedloopPolicy (local) and ActionChunkingClientSidePolicy (remote)
-    so chunking behavior is identical.
+class ActionChunkScheduler(ActionScheduler):
+    """Buffers one action chunk and replays it one step at a time.
+
+    Fetches a new chunk from the model only when the current one is exhausted.
+    Per-env tracking allows environments to refetch independently.
     """
 
     def __init__(
@@ -79,10 +81,14 @@ def get_action(self, fetch_chunk_fn: Callable[[], torch.Tensor]) -> torch.Tensor
 
         return action
 
-    def reset(self, env_ids: torch.Tensor | None = None) -> None:
+    def reset(self, env_ids: torch.Tensor | slice | None = None) -> None:
         """Reset chunking state for the given envs (all if None)."""
         if env_ids is None:
             env_ids = slice(None)
         self.current_action_chunk[env_ids] = 0.0
         self.current_action_index[env_ids] = -1
         self.env_requires_new_chunk[env_ids] = True
+
+
+# Backwards-compatibility alias
+ActionChunkingState = ActionChunkScheduler