Zhengyue2/pcm-vla-baseline

PCM-VLA: no-prior baseline

This checkpoint is the no-prior baseline of the PCM-VLA project.

Base model	lerobot/smolvla_base
Foundation prior	none (PCM frozen, alpha = 0)
Trainable params	lm_expert (98 M)
Training data	all 1693 episodes of lerobot/libero (4 suites)
Training steps	25 000, batch 16, lr 1e-4 cosine + 1000-step warmup
LIBERO-Spatial 10-task SR (50 ep / task)	4.8 %

The baseline is trained on the same data scale as the official lerobot/smolvla_libero reference (which reaches 65.6 %) so the comparison is fair on training-data quantity. The remaining gap reflects two design choices we deliberately do not change: keeping the 303 M-parameter VLM frozen, and reduced training compute. The point of this checkpoint is to establish that without a frozen 3D foundation model prior, the SmolVLA action expert cannot learn closed-loop-stable LIBERO manipulation from demos alone, even with 4× the data of our prior variants.

Use

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

python -m src.eval --ckpt ckpts/baseline/pcm_vla_state.pt \
    --task_ids "[0,1,2,3,4,5,6,7,8,9]" --n_episodes 50 \
    --output_dir results/baseline

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