OpenAI Gym environments and tools for training a 3D-printed quadruped robot (SpotMicro) using reinforcement learning.
Rex-gym is a collection of OpenAI Gym environments and training tools designed to train control policies for an open-source, 3D-printed quadruped robot called Rex. It uses reinforcement learning, specifically Proximal Policy Optimization (PPO), to teach the robot tasks like walking, galloping, and navigating uneven terrain in simulation, with the goal of transferring these policies to the physical robot. The project provides a framework for experimenting with hybrid control strategies that combine pre-defined trajectories with learned adaptations.
Robotics researchers, hobbyists, and developers interested in reinforcement learning for legged locomotion, particularly those working with quadruped robots and seeking to implement sim-to-real transfer. It's also suitable for educators and students exploring robotic control algorithms.
Developers choose Rex-gym because it offers a ready-to-use simulation environment tailored for a popular open-source robot platform (SpotMicro), with implemented PPO training and support for multiple terrains and tasks. Its hybrid policy approach provides flexibility in balancing manual control and autonomous learning, accelerating experimentation and deployment.
OpenAI Gym environments for an open-source quadruped robot (SpotMicro)
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Enables blending inverse kinematics with learned feedback via PPO, allowing control from fully manual to entirely autonomous, as detailed in the learning approach section.
Includes planes, random heightfields, hills, and mazes to test policy robustness, with visual GIF examples provided in the README for each terrain type.
Specifically designed for deploying trained policies to the physical Rex robot based on the open-source SpotMicro design, aiming for zero additional tuning per the project goal.
Supports galloping, walking, turning, and standing up with both forward and backward motion, offering a variety of pre-built Gym environments for quadruped training.
Several tasks in the README checklist, such as backward gallop with Bezier controller and maze navigation, are marked as not implemented, limiting immediate usability.
Tightly coupled to the Rex/SpotMicro robot platform; adapting to other quadrupeds requires significant modification of URDF models and kinematics, which is non-trivial.
Only implements PPO for reinforcement learning, with no built-in support for alternative algorithms, restricting experimentation without custom code changes.