PI0.5 PutCab Concurrent Train50 — E113, 10K updates
Inference checkpoint from experiment E113-R001, trained on 50 Concurrent demonstrations of the RoboTwin put_object_cabinet task. This is the Concurrent model, not the CTR-data model (E115).
Training
Initialized from the standard PI0.5 Base mirror XinY0201/openpi-pi05-base-jax at revision 5e62884fcf8cb8f9fc693c9163ea18d3e3739658, corresponding to the configured canonical initializer gs://openpi-assets/checkpoints/pi05_base/params. This Run did not resume an earlier task-finetuned checkpoint.
Dataset: Shiki42/PutCab-Concurrent-Train50-V4, 50 episodes and 11,038 frames. Batch16, 10,000 optimizer updates, seed87431, 160,000 sampled frames. Uses the recorded LoRA recipe; the vision backbone and action projections are also trainable under that recipe. See resolved_config.json for exact model and optimizer settings.
Evaluation
E114-R001 reported 25/100 successes on PutCab-FixedRole100-v1, with7 collision-positive episodes. Fixed roles: left drawer, right object. Evaluation:100Hz,700-action limit,50 denoising steps, fixed B4 inference. Results remain audit pending, not archived/confirmed. Success uses the RoboTwin task predicate, not a sustained collision-free placement test. This is one training seed on a fixed expert-qualified test suite.
Loading and contents
Contains Orbax JAX params/, dataset-specific normalization in assets/, a resolved configuration, provenance and SHA256SUMS. Use the compatible OpenPI checkout at e9ba7b7732a3e66e4bd87d6d3429f7cef6352ead, config pi05_putcab_athenb_fullhorizon_mb16_ga1_lora3ep, and load this repository directory as the checkpoint directory. The resolved configuration records portable asset paths; adapt those paths to your checkout.
Inputs are head/left-wrist/right-wrist RGB,14-dimensional robot state and task instruction. Output is a50-action chunk (14 robot action dimensions after transforms). Camera setup and robot embodiment must match the training/evaluation setup.
Optimizer state and data-loader state are excluded; this is not a resumable training checkpoint. A fresh-process reload passed before publication, and all uploaded parameter/asset files were independently rehashed against that receipt. Refer to upstream PI0.5/OpenPI and dataset terms for applicable usage conditions.