diff --git a/src/lerobot/scripts/lerobot_train.py b/src/lerobot/scripts/lerobot_train.py
index 286c69906..927e9659a 100644
--- a/src/lerobot/scripts/lerobot_train.py
+++ b/src/lerobot/scripts/lerobot_train.py
@@ -259,7 +259,14 @@ def train(cfg: TrainPipelineConfig, accelerator: Accelerator | None = None):
         from accelerate.utils import DistributedDataParallelKwargs
 
         ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
-        accelerator = Accelerator(step_scheduler_with_optimizer=False, kwargs_handlers=[ddp_kwargs])
+        # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
+        # Force the device to be CPU when policy.device is set to CPU.
+        force_cpu = cfg.policy.device == "cpu"
+        accelerator = Accelerator(
+            step_scheduler_with_optimizer=False,
+            kwargs_handlers=[ddp_kwargs],
+            cpu=force_cpu,
+        )
 
     init_logging(accelerator=accelerator)