hil-serl performs well during training, but performs poorly when using the eval_policy.py

[yfynb1111]

System Info

- `lerobot` version: 0.3.3
- Platform: Linux-6.8.0-83-generic-x86_64-with-glibc2.35
- Python version: 3.11.13
- Huggingface_hub version: 0.35.0
- Dataset version: 3.6.0
- Numpy version: 2.2.6
- PyTorch version (GPU?): 2.10.0.dev20250918+cu128 (True)
- Cuda version: 12080
- Using GPU in script?: <fill in>

Information

• One of the scripts in the examples/ folder of LeRobot
• My own task or dataset (give details below)

Reproduction

pick rate is almost 100% when traing
https://github.com/user-attachments/assets/0e84db06-aa08-4d93-a051-531fcb36afb9

but when eval（eval_policy.py）,pick rate is 0
https://github.com/user-attachments/assets/342871fd-5470-4a81-b2f6-0a5c83427fcd

Expected behavior

Recurrence solution

[yfynb1111]

sorry，The link above is invalid
when eval
https://github.com/user-attachments/assets/794ef2bc-5883-4b47-9790-2e148969c513
when training

t.mp4

[shs822]

how did you vis the sim environment while training? your version of lerobot is 0.3.3?

[yfynb1111]

I solved it when modified "policy.from_pretrained(env_cfg.pretrained_policy_name_or_path)" to "policy = policy.from_pretrained(env_cfg.pretrained_policy_name_or_path)"

[shs822]

I solved it when modified "policy.from_pretrained(env_cfg.pretrained_policy_name_or_path)" to "policy = policy.from_pretrained(env_cfg.pretrained_policy_name_or_path)"

newbee!!!!!