zzzjie-robot/leggedmanip_lab

LeggedManip Lab is a research framework for training legged robots (robots that walk on legs like a dog) that also have a robotic arm attached. The core challenge these robots present is that the legs and the arm must be controlled together: the robot needs to stay balanced while walking and simultaneously move its arm to reach, grab, or place objects. This project provides a unified training system for tackling that coordination problem. The training uses a technique called reinforcement learning, where a simulated robot learns by trial and error in a virtual environment. The simulation runs inside NVIDIA's Isaac Lab and Isaac Sim, which are physics simulation tools commonly used in robotics research. Once trained, a policy (the learned behavior) can be exported and tested inside a second, lighter simulation environment called MuJoCo, as a step toward eventually running it on a physical robot. The framework supports seven hardware combinations, each pairing a legged robot base from Unitree (such as Go1, Go2, B1, B2, or Aliengo) with a different robotic arm (such as a Unitree Z1, ARX-X5, or Agilex Piper). Each combination can be trained in two modes: Flat, which handles walking and arm movement on level ground, and Whole-Body Control (WBC), which tracks a target position and orientation for the arm's tip while the robot walks. After training, you can test a policy using keyboard controls: one set of keys moves the robot's body (forward, backward, left, right, turn), a second set moves the arm tip in space, and a third set rotates the arm tip. This lets researchers verify behavior interactively before attempting a real-world deployment. The project is in active development. Transferring policies to physical hardware (sim-to-real) is listed as coming soon. The codebase is written in Python and released under the Apache 2.0 license.