lxliam/legged_mpc_amp

This project is a data generation tool for four legged robots, built on nonlinear model predictive control and whole body control techniques. It produces AMP data, which is the kind of motion data used to train robots to imitate natural, animal-like walking patterns. The tool wraps keyboard control inside the Gazebo robot simulator, fully automatic recording of that motion data, and support for importing several different robot models. The project targets Ubuntu 20.04 with ROS Noetic. After installing a list of system dependencies with apt, you compile the whole workspace with a single setup script. To run it, you open two terminals: one launches the Gazebo simulation with an empty world, and the other launches keyboard control with AMP data logging turned on and pointed at an output folder. Once running, you press a key to initialize the controller, another key to switch to a trotting gait, and the w, s, a, and d keys to steer the robot forward, backward, and turn. Core features include one step automatic recording of one to two minutes of varied motion data per sample, direct export of that data into the .npz file format used by IsaacLab, a live plotting tool that shows the foot trajectories the model predictive controller calculates, and support for several existing quadruped robot models including Go1, Go2, A1, Aliengo, and Lite3, with a guide for adding new robots. Detailed documentation covers the build process, adding new robots, and the AMP data collection and conversion steps. The project builds on an existing open source legged robot control project and its own modifications are released under the BSD 3-Clause license.