Installation: DexJoCo¶
1. DexJoCo¶
Install DexJoCo and its dexjoco conda environment, then add
CleanRUA's requirements:
conda activate dexjoco
pip install 'httpx>=0.28,<1' 'pydantic>=2.11,<3' 'tyro>=0.9,<1' 'tensorboardX>=2.6' moviepy
Images render with EGL unless MUJOCO_GL is set.
2. Model API¶
A Chat Completions model with vision, tool calling and usage reporting (add --provider aliyun for Alibaba Cloud):
export CLEANRUA_MODEL_NAME=YOUR_VISION_MODEL
export CLEANRUA_BASE_URL=https://YOUR_PROVIDER/v1
export CLEANRUA_API_KEY=YOUR_API_KEY
3. Pi0.5¶
Every DexJoCo agent needs the multi-task Pi0.5 checkpoint from DexJoCo-Pi05,
served by OpenPI in DexJoCo's openpi environment. One checkpoint serves every task:
cd /path/to/dexjoco/openpi && bash install.bash && conda activate openpi
XLA_PYTHON_CLIENT_PREALLOCATE=false python scripts/serve_policy.py --port=6001 policy:checkpoint \
--policy.config=multi_task --policy.dir=/path/to/pi05_dexjoco_multi_task
It uses about 8.6 GB of GPU memory and listens on all interfaces, so keep port 6001 closed to other machines. To
serve it from another GPU host, run examples/pi05_server/serve_dexjoco.sh
there and forward the port with PORT=6001 LOCAL_PORT=6001 examples/pi05_server/tunnel.sh here.
4. Run¶
python -m cleanrua.delegate_dexjoco --native-root /path/to/dexjoco --task pick_bucket
python -m cleanrua.monitor_dexjoco --native-root /path/to/dexjoco --task pick_bucket
python -m cleanrua.code_delegate_dexjoco --native-root /path/to/dexjoco --task pick_bucket
--task takes any task in dexjoco/configs/multi_task, --seed picks the scene, and --policy-url defaults to
ws://localhost:6001. Runs are written as for RoboDojo, to runs/<task>__<exp_name>__<seed>__<time>/.