fix(robotwin): re-enable autograd for CuRobo planner warmup and take_action

lerobot_eval wraps the full rollout in torch.no_grad() (lerobot_eval.py:566),
but RoboTwin's setup_demo → load_robot → CuroboPlanner(...) runs
motion_gen.warmup(), which invokes Newton's-method trajectory optimization.
That optimizer calls cost.backward() internally, which raises

    RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

when autograd is disabled. take_action() hits the same planner path at every
step. Wrap both setup_demo and take_action in torch.enable_grad() so CuRobo's
optimizer can build its computation graph. Policy inference is unaffected —
rollout()'s inner torch.inference_mode() block around select_action() is
untouched, so we still don't allocate grad buffers during policy forward.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/lerobot/envs/robotwin.py

@@ -23,6 +23,7 @@ from typing import Any
 
 import gymnasium as gym
 import numpy as np
+import torch
 from gymnasium import spaces
 
 from lerobot.types import RobotObservation
@@ -320,6 +321,10 @@ class RoboTwinEnv(gym.Env):
         actual_seed = self.episode_index if seed is None else seed
         setup_kwargs = _load_robotwin_setup_kwargs(self.task_name)
         setup_kwargs.update(seed=actual_seed, is_test=True)
+        # setup_demo → load_robot → CuroboPlanner.warmup() runs Newton's-method
+        # trajectory optimization, which requires autograd. lerobot_eval wraps
+        # the whole rollout in torch.no_grad(), so re-enable grad here.
+        with torch.enable_grad():
             self._env.setup_demo(**setup_kwargs)
         self.episode_index += self._reset_stride
         self._step_count = 0
@@ -333,6 +338,9 @@ class RoboTwinEnv(gym.Env):
             raise ValueError(f"Expected 1-D action of shape ({ACTION_DIM},), got {action.shape}")
 
         # RoboTwin 2.0 uses take_action(); fall back to step() for older forks.
+        # take_action() invokes the CuRobo planner, which needs autograd —
+        # lerobot_eval wraps the rollout in torch.no_grad().
+        with torch.enable_grad():
             if hasattr(self._env, "take_action"):
                 self._env.take_action(action)
             else: