Fix eval action conversion for bf16 policies

src/lerobot/scripts/lerobot_eval.py

@@ -191,7 +191,7 @@ def rollout(
         action = action_transition[ACTION]
 
         # Convert to CPU / numpy.
-        action_numpy: np.ndarray = action.to("cpu").numpy()
+        action_numpy = _action_to_env_numpy(action)
         assert action_numpy.ndim == 2, "Action dimensions should be (batch, action_dim)"
 
         # Apply the next action.
@@ -261,6 +261,11 @@ def rollout(
     return ret
 
 
+def _action_to_env_numpy(action: Tensor) -> np.ndarray:
+    """Convert policy actions to a NumPy array accepted by Gym environments."""
+    return action.detach().to(device="cpu", dtype=torch.float32).numpy()
+
+
 def eval_policy(
     env: gym.vector.VectorEnv,
     policy: PreTrainedPolicy,

tests/scripts/test_lerobot_eval.py

@@ -0,0 +1,14 @@
+import numpy as np
+import torch
+
+from lerobot.scripts.lerobot_eval import _action_to_env_numpy
+
+
+def test_action_to_env_numpy_casts_bfloat16_to_float32():
+    action = torch.tensor([[0.5, -1.0]], dtype=torch.bfloat16)
+
+    action_numpy = _action_to_env_numpy(action)
+
+    assert action_numpy.shape == (1, 2)
+    assert action_numpy.dtype == np.float32
+    np.testing.assert_allclose(action_numpy, np.array([[0.5, -1.0]], dtype=np.float32))