# dexhand training kit (private) Everything a fresh GPU box that reaches only Hugging Face needs to train the dexhand pi0.5 policies, laid out as the workspace `$SYT` (download this repo into `$SYT`): - `code/XPolicyLab/policy/Pi_05/openpi/` - openpi as used (XPolicyLab Pi_05, with uv.lock; `uv sync` there) - `code/runner/` - training fixes the runner imports (gradient accumulation, state defragmentation, resume) - `code/dexhand/train/` - dexhand config (`dexhand_openpi.py`), entry points (`run.py`, `train.sh`), serving (`serve.py`) - `cache/openpi/` - openpi's cache: `openpi-assets/checkpoints/pi05_base/params` (pi0.5 base, JAX/orbax) and `big_vision/paligemma_tokenizer.model`; point `OPENPI_DATA_HOME` here - `env.sh.template` - the environment (set `SYT` first; drop the mirror lines if not in mainland China) - `dexhand_realrobot_kit.tar.gz` - real-robot kit (robot adapter interface, runner, replay, grip test, README) - `MANIFEST.sha256` - SHA-256 of every file above except the kit and this README Data: public HF dataset `Noctis12138/dexhand-midtrain` (`lerobot/_*seg_*`). pi0.5 base weights: Gemma terms, see `code/XPolicyLab/policy/Pi_05/openpi/LICENSE_GEMMA.txt`.