lukasskellijs/pi05_base_assembly_bench_2_1ep

pi0.5 base → assembly_bench_2 (1 epoch SFT)

Full fine-tune of lerobot/pi05_base (pure π₀.₅, not DROID-finetuned) on the full lukasskellijs/assembly_bench_2 demonstration set.

Training

Item	Value
Base	lerobot/pi05_base
Dataset	lukasskellijs/assembly_bench_2 (~1500 eps / ~508k frames)
Schedule	1 epoch · 16,953 steps
Batch	10 × grad_accum 3 (eff 30), bf16
LR	5e-5, warmup 847
Relative actions	yes (use_relative_actions=True)
Cameras	front→base_0_rgb, wrist→left_wrist_0_rgb (DROID keys)

Wandb: lgwlab/pi05-assembly-sft/17ssps3p

Eval (3-task × 8-env staged baseline)

Open-loop chunk_size=15 (HUD Adapter.chunk_size). Dense env total_reward:

Task	mean	best	lift
peg_round_M1_loose	0.201	0.928	1/8
gear_medium	0.238	1.139	1/8
nut_M16	0.154	0.171	0/8

HUD job: 5fad88b05f2a4f649153cb4b2d93b4f1

Load

from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("lukasskellijs/pi05_base_assembly_bench_2_1ep")

Note: checkpoint config declares observation.state shape 32 (max_state_dim), but assembly SFT norm stats are 8-D (7 joints + gripper). Feed 8-D proprio into the preprocessor (see assembly_bench pi05_assembly_sft agent).