backup wip

lerobot/scripts/eval.py

@@ -121,7 +121,7 @@ def eval_policy(
 
         # get the next action for the environment
         with torch.inference_mode():
-            action = policy.select_action(observation, step)
+            action = policy.select_action(observation, step=step)
 
         # apply inverse transform to unnormalize the action
         action = postprocess_action(action, transform)

lerobot/scripts/train.py

@@ -213,7 +213,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
         for key in batch:
             batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-        train_info = policy(batch, step)
+        train_info = policy.update(batch, step=step)
 
         # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.log_freq == 0: