Steps: 12000 (checkpoints every 1k; this repo is the final step)
Action space: DROID absolute joint-position (7 arm + gripper)
Load: PI05Policy.from_pretrained("lukasskellijs/pi05_assembly_bench_3_full_vlm_12k")
Train run: wandb

Intermediate step_00x000 snapshots are not included (ask if you need them).

lukasskellijs/pi05_assembly_bench_3_full_vlm_12k

pi0.5 full SFT on assembly_bench_3 (12k steps)