RLinf โ 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 ฯ0 rollouts (std-0.10 action head). These are the warm-start policies behind the WAN world-model PPO results in 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 |
the strong base for the WAN-PPO / pessimism results |
Both are plain PyTorch state_dicts that strict-load into the RLinf cnn_policy with
num_action_chunks=8 (image-only, actor_logstd = ln(0.10)).
Usage
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.