---
license: mit
tags:
  - robotics
  - libero
  - behavior-cloning
  - rlinf
---

# RLinf — LIBERO-spatial image-only chunk-8 CNN BC policies

Two pretrained (behavior-cloned) **image-only, action-chunk-8 CNN policies** for LIBERO-spatial
task 1, distilled from π0 rollouts (std-0.10 action head). These are the warm-start policies behind
the WAN world-model PPO results in [RLinf](https://github.com/JacobB33/RLinf) (stage-3 WM RL +
JEPA-consistency pessimism).

| file | strength (fixed 48-ep eval) | role |
| --- | --- | --- |
| `weak4_t1_chunk8_std010.pt` | ~0.40 (weak, 4 demos) | the weak base for the "does RL/pessimism help a weak policy" WAN-PPO experiments |
| `libero_spatial_t1_bc_pi0_imageonly_chunk8_std010.pt` | ~0.83 honest (~"76%") | the strong base for the WAN-PPO / pessimism results |

Both are plain PyTorch `state_dict`s that strict-load into the RLinf `cnn_policy` with
`num_action_chunks=8` (image-only, `actor_logstd` = ln(0.10)).

## Usage
```bash
hf download jacob3333/rlinf-libero-chunk8-cnn-bc --local-dir checkpoints/bc --include "*.pt"
```
The RLinf configs (`wan_libero_spatial_t1_ppo_cnn.yaml` and its sole/consistency variants) load the
strong one as `runner.ckpt_path`; the weak runs override with `weak4_t1_chunk8_std010.pt`.