scan-object-uniform — PI0.5 peer recipe LoRA10k — E169
Inference-only JAX/Orbax checkpoint from the authorized Sim12 training series. Pass 10000/ as checkpoint_dir. It contains full model parameters and paired normalization assets; optimizer, train_state, and data_loader resume state are excluded. No format conversion was performed.
Dataset: Shiki42/ctr-scan-object-uniform-20260911 at 000485f6b1cc3221f7ab21ddbda0d231757e8cf9. Training: 10000 optimizer updates, global batch16, GA1, FSDP1, seed87431, delta joint actions and temporal-padding loss mask. The exact qualified config and inference-facing OpenPI config are included under provenance. Set PARALLELVLA_DATASET_REPO and PARALLELVLA_NORM_ASSETS_DIR to this dataset and local 10000/assets; use the compatible PI0.5 OpenPI source and its required base-config environment. This is not a Transformers/safetensors model.
Action horizon50 is distinct from diffusion num_steps10. Evaluation results are tracked in https://swanlab.cn/@Travor/CTR-PI05-LoRA10k and are not implied by upload completion. Every source file was SHA-256 checked against the original CPU reload receipt before upload. CHECKPOINT_MANIFEST.json inventories inference files only.