policyweave-piper

PolicyWeave Piper inference weights

Inference checkpoints for the Piper branch of Daisy-zzz/PolicyWeave. manifest.json is the authoritative availability record. A pending entry does not indicate that usable weights have been uploaded.

Released components:

shared_base/
experts/adjust_bottle/
experts/stack_block/
experts/press_button/
experts/place_cup/
pi05/

Each GR00T expert contains only its verified10000-step inference adapter, configuration and shared normalization metadata. The Piper-adapted shared base is required and published separately. No optimizer states, raw demonstrations, training logs, or local machine paths are distributed.

Use scripts/download_piper_checkpoints.py on the code repository's piper branch (authenticate if needed). It checks availability and verifies every downloaded file against a SHA256 manifest pinned to a single repository revision.

The bank supports individual inference, dynamic_merge, WA, Core+TIES, Core+TSV and Core+ISO. The GR00T shared base is not cotraining.

pi0.5 four-task cotraining

pi05/ contains the final 30,000-step joint policy trained on51 episodes each of adjust_bottle, stack_block, press_button and place_cup. Training used physical batch256 without gradient accumulation, a frozen VLM, rank8/alpha16 LoRA on action attention/FFN, and full action input/output, time MLP and AdaRMSNorm training.

The self-contained release includes the converted original pi05 base, all334 trained parameter tensors, normalization, tokenizer, exact task prompts, inference config and SHA-256 manifest (approximately13.9GB). It does not use GR00T's shared base or specialist adapters. Separate LoRA and FP32 dense weights are retained; they are not rounded into the frozen BF16 base.

Use the pi0.5 deployment guide to set up its separate Python3.11/CUDA environment with OpenPI revision 215abfb217dbac7d5f1273282331b9b1866c0479 and its Transformers4.53.2 patch:

python scripts/download_piper_checkpoints.py --component pi05
python scripts/smoke_pi05_piper.py --checkpoint checkpoints/piper/pi05
python -m policyweave.piper_server --method pi05 \
  --checkpoint checkpoints/piper/pi05 --denoising-steps 10

pi0.5 uses two RGB cameras with PIL aspect-preserving resize/padding to224×224, the exact training state-token and q01/q99 normalization rules, and10 Euler steps by default. Outputs are16×7 absolute joint/gripper targets in radians/meters. It uses the existing single-arm HTTP robot client. No optimizer state, demonstrations or training-machine paths are included.

Read the deployment guide before use. Single-arm absolute7D joint/gripper actions require verified camera mapping, units and physical limits. Default execution is dry-run. These weights are experimental and are not certified for unsupervised robot operation.

Underlying GR00T, pi0.5 and tokenizer artifacts remain subject to their upstream terms; the code repository's software license does not replace those terms.