Fix eval action conversion for bf16 policies

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))