pi05-bread-r1-top200 — RA-BC policy for bread-to-toaster (bimanual UMI)

pi0.5 base + LoRA finetune (10k steps, killed at loss plateau), trained with RA-BC (reward-model-weighted BC, SARM2 delta-progress weights, clipped at 0) on 200 curated bimanual UMI episodes (top-200 of 256 by reward-model mean-dP + dP-variance).

Training code: https://github.com/BrianZhengJourney/openpi/tree/bread-r1 (pi05_bread_r1_lora config; BreadInputs/BreadOutputs in src/openpi/policies/bread_policy.py). Data pipeline + decision docs: https://github.com/BrianZhengJourney/bread-toaster.

Contents

• 10000/params — orbax params (servable with openpi create_trained_policy)
• 10000/assets — norm stats (quantile), computed on the 200 training episodes. The policy is meaningless without them; loaded automatically from the ckpt dir.

Inference contract (must match EXACTLY at deployment)

Item	Value
Obs images	left_wrist_image, right_wrist_image — uint8 HWC, resize-with-pad → 224×224 (done inside the transform stack). No base/top camera: the base_0_rgb slot is zeroed + masked (training top view was a human egocam; do not feed the RealSense)
Obs state	20D [Lxyz, Lrot6d, Lgrip, Rxyz, Rrot6d, Rgrip], relative_to_first: per arm T_rel = inv(T_arm,start) @ T_arm,t (record each arm's grasp_site pose at rollout start, re-express every subsequent pose). Frame = URDF grasp_site (yam_linear_4310 canonical TCP). Gripper: 0=closed, 1=open
Prompt	take the bread out of the bowl and put the bread into the toaster (exact string)
Action	(40, 14) @30 Hz: per arm [Δxyz(3, TCP frame), axis-angle Δrot(3), ABSOLUTE gripper target(1)], stepwise-chained: delta_k = inv(T_{t+k-1}) @ T_{t+k}; executor integrates T_cmd,k = T_cmd,k-1 @ delta_k starting from T_current. SO(3) composition only; denormalize exactly once
Latency	UMI-style: feed the proprio sampled at image exposure time; start executing the chunk at step ceil(latency × 30 Hz)

Deployment (recommended: client/server over LAN)

Robot NUC runs only the pure-CPU openpi_client; the model serves from a lab GPU workstation (>=16GB VRAM). See deploy/serve_bread.sh (server, one command) and deploy/client_example.py (the fixed obs-in -> actions-out contract for the executor).

Minimal loading example (openpi, branch bread-r1)

from openpi.training import config as _config
from openpi.policies import policy_config

policy = policy_config.create_trained_policy(
    _config.get_config("pi05_bread_r1_lora"), "<this-repo-local-path>/10000")
out = policy.infer({
    "left_wrist_image": left_uint8_hwc, "right_wrist_image": right_uint8_hwc,
    "state": rel_state_20d,
    "prompt": "take the bread out of the bowl and put the bread into the toaster",
})
actions = out["actions"][:, :14]   # (40, 14)

Provenance

• Base: gs://openpi-assets/checkpoints/pi05_base · LoRA gemma_2b_lora + gemma_300m_lora · bs 32 · lr 2.5e-5 cosine · stopped at 10k/20k (plateau; lr was ~50% peak — resume from 10000 if undertrained on-robot)
• Reward model for weights: SARM2 r4_bimanual (v3 7-stage labels, human-verified boundaries)
• wandb: project openpi, run r1_top200 (8d6mivlw)
• Offline P0 check results: see the repo issue/notes accompanying this upload