Msornerrrr / in-hand-rotation-mjlab

What is in-hand-rotation-mjlab?

This Python repo delivers a complete sim-to-real transfer GitHub pipeline for in-hand cube rotation using the LEAP hand, built on Mjlab and MuJoCo. It handles training RL policies with domain randomization, asymmetric actor-critic observations (noisy history for actor, full state for critic), and curriculum rewards, then deploys them to hardware via a ZMQ policy server at 20Hz. Developers get parallel training (up to 4096 envs), sim2sim testing, grasp cache generation, and Dockerized ROS tools for real-robot sysid and control.

Why is it gaining traction?

It stands out with end-to-end sim-to-real transfer of accurate grasping via eye-in-hand observations and continuous control, skipping fragmented setups common in manipulation research. Key hooks include pre-trained checkpoints for instant play, trajectory replay servers, and actuator calibration from real logs—making validation fast without custom glue code. The modular Mjlab base enables quick tweaks to hands, cubes, or rewards.

Who should use this?

Robotics researchers prototyping dexterous in-hand manipulation on LEAP hands or similar underactuated designs. Ideal for teams bridging sim policies to real hardware, like those tuning cube rotation or grasp stability with eye-in-hand cams. Skip if you're not in MuJoCo RL or need production-grade reliability.

Verdict

Solid research prototype for sim-to-real GitHub workflows—grab it if in-hand hand rotation fits your stack, with strong docs and example policies easing starts. Low 19 stars and 1.0% credibility score signal early maturity; test thoroughly before production, but fork and contribute to mature it.

(198 words)