nori-a3-mujoco

Unofficial MuJoCo model of the Nori A3. It uses the session API from the Nori Python SDK. Replace mock_session() with mujoco_session(MujocoRobot()). The rest of the SDK code stays.

Nori Robotics has not reviewed this. Kinematics come from nori_description v0.1.1. Mass, servo strength, and calibration are guesses. Read "What is approximate" before you trust a number.

from nori_a3_mujoco import MujocoRobot, mujoco_session

async with mujoco_session(MujocoRobot()) as robot:
    ...

mujoco_session runs in realtime and records a video track. Pass realtime=False or video=False to turn either one off.

Install on Arch

You need Python 3.11, 3.12, or 3.13. .python-version pins 3.12. Install uv.

sudo pacman -S --needed uv mesa ffmpeg
sudo pacman -S --needed intel-compute-runtime level-zero-loader
make setup-xpu

make setup-xpu installs torch==2.9.1+xpu from the PyTorch XPU index. The wheels include the oneAPI runtime. make setup installs the CPU torch build instead.

make setup copies .env.example to .env. The Makefile exports that file.

The pin is nori-sdk 1.1.0, from nori-sdk-py at a8db25b.

If you do not use uv, install requirements/xpu.txt and requirements/dev.txt, then run pip install -e . --no-deps.

Run a demo

make test
make example
make demo
make view
make gpu-info

make test runs 40 tests with no window. The test that diffs sdk_mock_pick_place.py against the SDK skips unless ../nori-sdk-py is checked out beside this repo.

make example runs the pick-and-place with no window. make demo writes videos/demo.mp4 and stills. make view opens a GLFW window. Drag a joint with the mouse.

Offscreen frames use MUJOCO_GL=egl. Set NORI_VIEWER=1 to attach the GLFW viewer to every session.

With Mesa 26.2.2 on an Arc B580, offscreen EGL used the Arc. make demo wrote 410 frames, 13.7 s of video, in 55 s of wall time. That is about 0.13 s per frame. Physics stays on the CPU. The examples time their waits on the sim clock, because a recording runs slower than the video.

Drive it from SDK code

from nori_sdk.motion import JogBuilder
from nori_a3_mujoco import MujocoRobot, mujoco_session

async with mujoco_session(MujocoRobot()) as robot:
    info = await robot.wait_ready()
    await robot.jog(JogBuilder(info.descriptor).base(linear=0.4).build(), duration=1.0)
    await robot.action({"left_arm_gripper.pos": 30.0}, wait=True)
    frames = await robot.snapshot()
    robot.estop()

action and pose finish from the physics. A joint that settles inside the tolerance ends as done. A joint that hits a limit ends as clamped. estop latches until reset_latch.

run_policy(robot, fn, hz) sends action, jog, and pose at a fixed rate. The same function accepts this sim, mock_session(), or a real RemoteTeleop. If fn stops answering, the runner lets the SDK watchdog trip. Return STOP to end the run.

NoriA3Env in gym_env.py wraps the same handler for Gymnasium. It has no task reward.

Run a torch policy on an Arc

NORI_DEVICE=auto picks CUDA, then XPU, then CPU. Set NORI_DEVICE to xpu, xpu:0, cuda, or cpu. If that device is missing, the code warns and uses the CPU.

NORI_DEVICE=xpu uv run python examples/torch_policy.py --bench 200

The bench warms up, then times the forward pass and the copy back to the host. On an Arc B580 with torch 2.9.1+xpu, 200 calls took a median of 1.60 ms, 95th percentile 2.10 ms. The same wheel on CPU took a median of 5.87 ms, 95th percentile 7.04 ms.

If an integrated GPU is also visible, set ONEAPI_DEVICE_SELECTOR=level_zero:0.

What is approximate

From the URDF and the SDK, unchanged:

• Kinematic tree, joint axes, limits, visual meshes, and camera mounts.
• Actuator names, units, camera roles, the estop latch, and the watchdog profiles.
• examples/sdk_mock_pick_place.py matches the SDK file except the import and the session constructor.

Guessed, and not measured on a robot:

• Link masses sum to about 18.3 kg. The robot is about 20.5 kg. This model does not rescale them.
• Arm joints are position servos with kp 300 and kv 4. The torque cap on each joint is a guessed split.
• SDK units map in a straight line onto the URDF limits. Gripper 0 is closed and 100 is open. The real calibration offsets are unknown.
• Wheel radius is 0.0762 m. Track is 0.30 m.
• The descriptor allows a 720 mm lift. The model stops at 700 mm.
• The left claw can pick the cube up by friction in one window. Put the cube about 32 to 36 cm forward of the base and 8 to 20 cm to the side. Set the lift to 300 mm. It lands 6 to 8 mm from the marker. The pad shape and the grip force are ours.
• The left_wrist picture is upside down beside right_wrist. It stays that way.

Navigation goals, episode recording, and policy_stream still take the mock path. This sim does not run them.

Rebuild the model

git clone --branch v0.1.1 https://github.com/Nori-Robotics/nori_description ../nori_description
make build-model DESCRIPTION=../nori_description

scripts/build_model.py writes assets/a3.xml and the meshes. Do not edit a3.xml by hand.

License

Code in src/, scripts/, examples/, and tests/ is Apache-2.0. See LICENSE. examples/sim_pick_place.py comes from the Nori SDK example.

assets/a3.xml and the meshes come from nori_description v0.1.1, copyright Nori Robotics Inc., under CC BY-NC-SA 4.0. See assets/UPSTREAM_LICENSE. Non-commercial, and share-alike. Nori calls that description source-available, not open source.