Skip to main content
wuji-mjlab is an in-hand cube reorientation reinforcement learning project for the Wuji Hand. It trains PPO policies in mjlab that cover the full SO(3) goal space, then deploys them on the physical hand through a sim-to-real bridge. The project ships pretrained checkpoints, deployment scripts, and hardware guides so you can reproduce the demo end to end.
This release targets the Wuji Hand. The deploy bridge wraps the Wuji Hand SDK and expects the matching firmware revision.
Simulation reorient demoReal-hand reorient demo

​
Task

The project centers on one task, with two training configurations. The policy holds a cube in a downward-facing hand, receives a target orientation in the palm frame, and rotates the cube in place until its orientation matches the goal within a hold window, without dropping it.

​
What You Get

  • A calibrated pipeline that runs the trained ONNX policy on real Wuji Hand hardware
  • Pretrained checkpoints and released policy assets, so you can deploy without training
  • A 3D-printed ArUco-tagged cube, wrist AprilTag world frame, and camera calibration workflow
  • Two training configurations: a release config that reproduces the demo, and a lower-VRAM variant

​
Get the Released Assets

Pull the pretrained checkpoint and CAD bundle from the latest release:
Without the gh CLI, open the latest release page, download the wuji-mjlab-v*-assets.zip attachment by hand, then run the same unzip and mv pair.