vla_jepa_so101_block_stack_sim

VLA-JEPA fine-tuned to stack one 30 mm block on another with an SO-101, in simulation. Trained by train/vla_jepa_so101.yaml in vla-test on drwahl/so101_block_stack_sim.

LoRA (r=32, alpha 64) on the Qwen language backbone of lerobot/VLA-JEPA-Pretrain, with the action model and the video predictor dense, vision 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/vla_jepa_so101_block_stack_sim --local-dir vla_jepa_so101_block_stack_sim
uv run python sim/scripts/eval.py vla_jepa_so101_block_stack_sim/adapter --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 7-step chunk, the setting that stacks most on one seed, it stacks 91.7 of 150 and lifts 105.7, averaged over three further seeds. A 5-step horizon stacks 59 on one seed against 93 for the whole chunk.