PI0.5 PutCab CTR Train50 — peer recipe — E149
Verified inference-only OpenPI JAX/Orbax checkpoint after 10000 optimizer updates. Pass 10000/ as checkpoint_dir; includes parameters and normalization assets only; optimizer/train_state and data_loader resume state are intentionally excluded. This is NOT a Hugging Face Transformers or safetensors checkpoint.
Fresh training from PI0.5 base revision 5e62884fcf8cb8f9fc693c9163ea18d3e3739658. Exact OpenPI source e9ba7b7732a3e66e4bd87d6d3429f7cef6352ead is supplied as a git bundle under provenance. Config pi05_putcab_athenb_fullhorizon_mb16_ga1_lora3ep, actual CLI10000updates overrides its historical default. Dataset Shiki42/PutCab-CTR-Train50-V4 at 3b9629a729945005dabd7857803218d063134634. Two H100 GPUs; global batch16, perGPU8, GA1, fsdp1, seed87431.
Use the supplied source and this checkpoint's own assets. Set PARALLELVLA_DATASET_REPO to the dataset above, PARALLELVLA_NORM_ASSETS_DIR to the local 10000/assets, and PI05_JAX_BASE to your verified base params location when importing the configuration. Joint actions use delta conversion, grippers remain absolute; adapt_to_pi=False. Temporal padding is masked in training, not idle-arm masking. Action horizon50 is distinct from inference num_steps10. Saved absolute source paths in provenance are historical, not portable defaults.
PutCab-FixedRole100-v1 / RT4 SPP / depth4 / OptiX / actual diffusion10: 63/100 successful. Frozen suite retains a documented75 prompt/asset category mismatch, so this score is specific to that protocol. Fixed-wave and cross-wave batching produced identical per-seed action hashes in the separate comparison. Reports are computationally verified, not claimed externally audit-approved.
All checkpoint source SHA256 values were rechecked before upload; original fresh-process CPU reload found all model parameters finite. CHECKPOINT_MANIFEST.json enumerates original checkpoint files. No model conversion was performed. This repository is public as explicitly requested by the owner.