pi0_so101_block_stack_sim

pi0 fine-tuned to stack one 30 mm block on another with an SO-101, in simulation. Trained by train/pi0_so101.yaml in vla-test on drwahl/so101_block_stack_sim.

LoRA (r=32, alpha 64) on the gemma_2b language backbone of lerobot/pi0_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/pi0_so101_block_stack_sim --local-dir pi0_so101_block_stack_sim
uv run python sim/scripts/eval.py pi0_so101_block_stack_sim/adapter --no-rtc --no-smooth --horizon 10

The eval rolls the policy out on the 150 held-out layouts and scores it by the gates that screened the oracle. Executing 10 steps of each 50-step chunk, the setting that stacks most on one seed, it stacks 114.7 of 150 and lifts 132.7, averaged over three further seeds. Executed whole, its chunk stacks 48.7. Stitching does not help it: at 20 steps it costs 20 layouts.