You can deploy this model and start earning money today!

Zhengyue2/
pcm-vla-raft

Name: pcm-vla-raft
Availability: InStock

Views

No views yet

PCM-VLA: RAFT-Small flow prior

This checkpoint is part of the PCM-VLA project: a small (~1 M-param) FiLM module that lets a frozen 3D foundation model modulate the SmolVLA action expert.


Base model	`lerobot/smolvla_base`
Foundation prior	RAFT-Small (1 M, frozen)
Trainable params	lm_expert (98 M) + 1 M PCM
Training data	432 episodes from `lerobot/libero` (libero_spatial subset)
Training steps	15 000, batch 16, lr 1e-4 cosine + 500-step warmup
LIBERO-Spatial 10-task SR (50 ep / task)	19.6 %

Use

bash

pip install huggingface_hub
python - <<PY
from huggingface_hub import hf_hub_download
hf_hub_download("Zhengyue2/pcm-vla-raft", "pcm_vla_state.pt", local_dir="ckpts/raft")
PY

# Then with the PCM-VLA codebase:
python -m src.eval --ckpt ckpts/raft/pcm_vla_state.pt \
    --task_ids "[0,1,2,3,4,5,6,7,8,9]" --n_episodes 50 \
    --output_dir results/raft

See github.com/ZhengyueZhao/pcm-vla for the full codebase, training instructions, and per-task results.