PI0.5 scan-object-concurrent — refreshed 20260916 data — E195
Inference-only OpenPI JAX/Orbax checkpoint. Use 10000/ as checkpoint_dir.
Full inference parameters and matching normalization assets are included. Optimizer, data-loader resume state and training-state checkpoints are excluded; no safetensors conversion was performed.
Dataset revision: 9c9483af251e9076bab5c31aff97628bf41e1764. Training: fresh base, 10000 optimizer updates, batch16 per experiment, GA1, FSDP1, seed87431, action horizon50, 12 joint dimensions represented as delta and grippers absolute, temporal-padding loss mask. Two independent experiments shared each H100. The standard OpenPI LoRA freeze rule also trains the vision encoder and heads.
Set PARALLELVLA_DATASET_REPO=Shiki42/ctr-scan-object-concurrent-20260916 and PARALLELVLA_NORM_ASSETS_DIR to your local 10000/assets. Use compatible OpenPI source commit e0e08ba202d53c9da05bd241eeb9c177c2b45e0d with config pi05_putcab_athenb_fullhorizon_mb16_ga1_lora3ep; its historical name does not change the dataset/task above. Relevant config/transform reference files are included in source_reference; this is not a standalone Python installation. PI05_JAX_BASE must point to a separately verified PI0.5 base parameters directory when required by that config.
The source checkpoint passed an independent CPU reload and finite-parameter check. Source inference files were rehashed before upload; see CHECKPOINT_MANIFEST.json. Training metrics are in the existing E195 run of https://swanlab.cn/@Travor/CTR-PI05-LoRA10k . No simulation success rate is claimed for this new checkpoint. This repo is separate from the earlier E168 model.