pi05-piperx-workpiece-storage-62ep-20k

PI0.5 workpiece storage, 62 episodes, full fine-tuning, final step 20000

Initialized from original PI0.5 base b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba, not a previously fine-tuned model. Dataset: Shiki42/piperx-workpiece-storage-0909-62ep-raw, immutable revision 5a924d4b954def471c5f57b7896886057788fc8a. All 62 episodes, 53366 frames at 30 FPS retained. Three RGB cameras: right_environment_1 (base), left_wrist, right_wrist. State/action:14 dimensions. Raw parquet preserved byte-for-byte. Action/state quantiles freshly computed over all frames using numpy.quantile linear; training stats SHA256 1ef7f492a09b93a1bda5264a7dec03a2c30692a7c289e15bcd9ceaa1a42fa0a7.

Task: put the copper screw into the left box and the black sleeves into the right box.

8 H100, FP32, global batch16 (2 per rank), horizon50, seed1000, 20000 updates, save5000. Full fine-tuning, no idle mask, no LoRA, no frozen vision encoder or expert-only training. LeRobot0.6 source commit1b8278b8fa8aa1e7c5430d3aa022102a646fbbc5, transformers5.3.0. Training ended2026-09-09 23:23:53 CST; checkpoint20000 validated selection_loss0.004833; processExitCode0.

The model weight and normalization tensors are preserved byte-for-byte. Deployment config metadata has been made portable; originals are retained in runtime/. CHECKPOINT_MANIFEST.json records SHA256 and sizes. No optimizer/scheduler/RNG states, dataset or intermediate checkpoints are included.

Loading: use the matching existing LeRobot environment. Load weights from the downloaded checkpoint directory instead of the saved platform-specific pretrained_path. Override tokenizer_processor.tokenizer_name with an existing local PaliGemma tokenizer (google/paligemma-3b-pt-224 revision35e4f46485b4d07967e7e9935bc3786aad50687c), and device_processor.device with the intended device. Map observation.images.right_environment_1 to observation.images.base_0_rgb and wrists to the corresponding *_0_rgb names. Saved /inspire paths are provenance paths, not portable local paths. This repository does not configure or actuate hardware.

EvoMind runtime metadata

evomind-runtime.json is next to config.json. See runtime/README.md for the distinction between the original NGC training environment and the existing validated PI0.5 inference environment, exact captured package inventory, tokenizer provenance, and receiver binding. FP32/eager/sync, AMP and compile disabled. Configuration/schema/interpreter/processor checks passed; this checkpoint has not undergone a new inference benchmark or robot validation. Set EVOMIND_PYTHON to the matching existing executable (or create runtime/environment-path.local) and chmod +x runtime/bin/python after download. No machine-local binding is distributed.