fix(peft): fix LoRA resume from Hub (PosixPath + double wrap) (#3485)

src/lerobot/policies/factory.py

@@ -514,7 +514,7 @@ def make_policy(
 
         logging.info("Loading policy's PEFT adapter.")
 
-        peft_pretrained_path = cfg.pretrained_path
+        peft_pretrained_path = str(cfg.pretrained_path)
         peft_config = PeftConfig.from_pretrained(peft_pretrained_path)
 
         kwargs["pretrained_name_or_path"] = peft_config.base_model_name_or_path
@@ -527,7 +527,9 @@ def make_policy(
             )
 
         policy = policy_cls.from_pretrained(**kwargs)
-        policy = PeftModel.from_pretrained(policy, peft_pretrained_path, config=peft_config)
+        policy = PeftModel.from_pretrained(
+            policy, peft_pretrained_path, config=peft_config, is_trainable=True
+        )
 
     else:
         # Make a fresh policy.

src/lerobot/scripts/lerobot_train.py

@@ -277,6 +277,11 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     if cfg.peft is not None:
         if cfg.is_reward_model_training:
             raise ValueError("PEFT is only supported for policy training. ")
+        from peft import PeftModel
+
+        if isinstance(policy, PeftModel):
+            logging.info("PEFT adapter already loaded from checkpoint, skipping wrap_with_peft.")
+        else:
             logging.info("Using PEFT! Wrapping model.")
             peft_cli_overrides = dataclasses.asdict(cfg.peft)
             policy = policy.wrap_with_peft(peft_cli_overrides=peft_cli_overrides)