smolvla_so101_block_stack_sim

SmolVLA fine-tuned to stack one 30 mm block on another with an SO-101, in simulation. Trained by train/smolvla_so101.yaml in vla-test on drwahl/so101_block_stack_sim.

LoRA (r=32, alpha 64) on the SmolVLM2 language backbone of lerobot/smolvla_base, with the action expert and the projection heads dense, SigLIP frozen: 20k steps at batch 4, the protocol drwahl/pi05_so101_block_stack_sim was trained on. Normalization, learning rate and chunk length are the base checkpoint's own. adapter/ holds the adapter and the processors, which carry the normalization and the camera renaming; lerobot loads it onto the base weights.

hf download drwahl/smolvla_so101_block_stack_sim --local-dir smolvla_so101_block_stack_sim
uv run python sim/scripts/eval.py smolvla_so101_block_stack_sim/adapter --no-rtc --no-smooth

The eval rolls the policy out on the 150 held-out layouts and scores it by the gates that screened the oracle. Executing its whole 50-step chunk, the setting that stacks most on one seed, it stacks 0.3 of 150 and lifts 11.3, averaged over three seeds. It predicts the demonstrations' actions on their own frames about as closely as the other policies do, and does not recover from its own errors once it drives the arm.