axiboai/pi05_piper_stacking_h25

π₀.₅ · piper_stacking · action_horizon 25 — FINAL (step 10,000)

π₀.₅ policy fine-tuned with openpi on the bimanual PiperX cube-stacking task. action_horizon=25 (0.83 s/chunk @30 fps) for more reactive control. This is the final 10k checkpoint at the repo root (replaces the earlier step-2,500 model, still in git history).

Base pi05_base · full fine-tune (no LoRA) · batch 32 · final train loss 0.0017. Embodiment piperx_bimanual (14-D state/action, cams cam_front/cam_left_wrist/cam_right_wrist). Contents: params/ (EMA inference weights) + assets/piperx_bimanual/norm_stats.json + _CHECKPOINT_METADATA (29 GB optimizer train_state/ excluded).

Load (openpi)

Requires the piperx-openpi fork with the pi05_piper_stacking_h25 config (see PR #11).

import pathlib
from huggingface_hub import snapshot_download
from openpi.training import config as _config
from openpi.policies import policy_config
ckpt = pathlib.Path(snapshot_download("axiboai/pi05_piper_stacking_h25"))
policy = policy_config.create_trained_policy(_config.get_config("pi05_piper_stacking_h25"), ckpt)
action = policy.infer(obs)["actions"]