decision-robot-control

d1 names the skill. MicroDuck's policies walk, turn, pick, and sit.

Six boxes. The white disc is the next gap. The model reads the picture and names the skill. The policies move the joints.

Run

uv sync
cp .env.example .env

DECISION_BASE_URL is a local llama.cpp server, or any host that accepts POST /v1/systemone. Set DECISION_MODEL and DECISION_API_KEY when the host asks for them. DECISION_URL is the old name for the base URL.

llama-server -m d1-3B-Q8_0.gguf --mmproj mmproj-d1-3B-Q8_0.gguf --port 8091 --host 127.0.0.1 -c 2048

CUDA build on NVIDIA. SYCL build on Arc. Same flags.

git clone --depth 1 https://github.com/pollen-robotics/microduck_rl

Set MICRODUCK_SCENE to scene.xml in that clone. The meshes stay there. The three policies in policies/ are Pollen's, Apache-2.0.

MUJOCO_GL=egl uv run python packages/microduck-harness/full_demo.py --vision

--vision leaves distance and bearing out of the request.

Policies are ONNX. onnxruntime-gpu uses CUDA. onnxruntime-openvino uses an Intel GPU. ONNX Runtime has no XPU provider. Without one of those wheels the policies stay on CPU.

The scene's position actuators track a medium walk command more slowly than the command says.