ruoyiqiao/mjlab_hand

This repository contains simulation environments for training robotic hands to grasp objects and spin them in place. It is built on top of mjlab, a framework for running robot-learning experiments on GPU hardware. The hands it supports are Allegro, LEAP, Shadow, Sharpa, and Wuji, each of which is a different design of multi-fingered robotic hand used in research. The two core tasks are grasping, where the hand must pick up an object without dropping it, and in-hand rotation, where the hand must keep spinning an object continuously. Training is done by running many simulated copies of the task at once, thousands of environments at a time, to collect experience quickly. You point it at a task name and a hand model, set how many parallel environments to run, and let it train for a set number of iterations. The repository also implements a method called ContactExplorer. It adds two extra reward signals on top of the basic task reward. One reward encourages the hand to make contact with parts of the object it has not touched before. The other guides the fingers toward under-explored regions of the object before contact is made. Together these signals are meant to help the hand discover a wider variety of contact strategies rather than converging on one narrow grip early in training. You can swap in custom objects from a dataset called ContactDB, which includes shapes like cups, mugs, hammers, and game controllers. Not all objects have been tested, and the README notes that some may need extra tuning to learn successfully. Once training is done, you can watch a trained policy interact with objects in an interactive viewer or run batch evaluation that prints a success-rate metric and optionally logs results to Weights and Biases. Installation requires Python 3.13 or later and uses the uv package manager. The robot hardware models come from publicly available collections of robot description files.