base/README.md

metadata

license: cc-by-nc-4.0
tags:
  - robotics
  - vla
  - pi05
  - openpi
  - manipulation

base

Ordinary pi0.5 conditioning: one frame, absolute proprioceptive state. The reference the other two are measured against.

training config	pi05_frame_base
checkpoint step	13,249 of 13,250
samples seen	423,968
batch size	32
peak LR	0.0005, cosine over 13,250 steps
conditioning	1 frame(s), state absolute
action space	relative EE, world_rotvec, horizon 50
base weights	pi05_base (PyTorch), LoRA r16 all_linear
frame yaw aug	none
normalizer key	mix/tuned_h50

Contents

• model.safetensors — the weights
• norm_stats.json — the quantile normalizer this checkpoint was trained against. Not interchangeable between variants: the absolute-pose and relative-motion state representations have different scales, and using the wrong one silently rescales output.
• train_config.json — the resolved training config
• mixture.json — dataset sampling weights

Loading

from experiments.infer import VariantPolicy
policy = VariantPolicy.load("base")
actions = policy.infer(images, poses, prompt)   # absolute EE targets, [horizon, 10]

Provenance

Trained from pi05_base (Physical Intelligence openpi) on cong-lab manipulation data plus a DROID subset. Inherits the licence terms of both.

Caveats

Trained on a 11-hour single-GPU budget, which is roughly 0.10 of one pass over the 4,245,385-sample mixture. These are ablation checkpoints for a specific comparison, not a converged production policy.