| # π0.5 cloth-folding inference checkpoint |
|
|
| ## Released files |
|
|
| ```text |
| checkpoints/pi05_fold_clothes/epoch_2/ |
| model.safetensors |
| metadata.pt |
| assets/umi_v3/norm_stats.json |
| ``` |
|
|
| This is epoch 2 of `pi05_umi_v3`. `model.safetensors` is 7,473,091,464 bytes, SHA256 `3e5d551d8cbda25b84fb2fcdb610e2936fbe7668c640ec3cbabd9e9f1ba9540f`. Tensor bytes and `norm_stats.json` are unchanged and were checked against OSS CRC64 after download. |
|
|
| `metadata.pt` was rewritten as a plain dictionary containing model/data inference settings, config name, epoch and global step. Training-storage paths, dataset roots, experiment names and training-only objects/settings were removed. The quaternion advice reader uses `torch.load(..., weights_only=True)`. The generic model loader uses the explicit `pi05_umi_v3` config and checkpoint normalization statistics; metadata is not used to restore training state. `optimizer.pt` is unnecessary for inference and is not included. |
|
|
| ## Load with the robot adapter |
|
|
| After installing the OpenPI environment described in the root README: |
|
|
| ```bash |
| export PI05_PYTHON=/absolute/path/to/openpi-umi/.venv/bin/python |
| export TOKENIZER_DIR=/absolute/path/to/tokenizer_root |
| export UMI_INFER_DATA=/absolute/path/to/calibrated_cell |
| bash scripts/run_pi05_server.sh |
| ``` |
|
|
| The launcher defaults to `checkpoints/pi05_fold_clothes/epoch_2/`, selects `pi05_umi_v3`, creates the local transformers 4.53.2 shadow, and automatically loads `assets/umi_v3/norm_stats.json`. The offline tokenizer must be at `TOKENIZER_DIR/paligemma-3b-pt-224/tokenizer.model`; tokenizer assets and device calibration are separate requirements. |
|
|
| The saved model config matches the bundled inference config: π0.5, 32 padded action dimensions, horizon 50, 200 tokens, Gemma 2B / 300M action expert, and BF16. Data uses both wrist images, wxyz pose labels, rot6d model features and no inter-gripper state. Keep `VLA_QUAT_PERM=0`. The adapter accepts previous/current 16-D poses, restores per-chunk anchors and decodes robot targets with the configured calibration. |
|
|
| The server defaults to 10 denoising steps (`PI05_DENOISE_STEPS`). Use the checkpoint's own statistics; do not recompute them from a demonstration sample. |
|
|
| ## Validation |
|
|
| Download integrity, safetensors structure, metadata sanitization and model/data config wiring were checked on CPU. CUDA inference and physical-robot cloth folding were not tested in this release. See [VALIDATION.md](VALIDATION.md). |
|
|
