lupinjia/LeggedGym-Ex

What it solves

LeggedGym-Ex provides a unified framework for training legged robots using reinforcement learning (RL), allowing researchers to easily switch between different physics simulators and implement various state-of-the-art locomotion methods without rewriting their entire pipeline.

How it works

Built upon the legged_gym architecture, the framework integrates three major simulators: IsaacGym (for high-speed training), Genesis (for speed and support of soft materials/fluids), and IsaacSim (for high-fidelity rendering). It uses Nvidia Warp to accelerate depth image rendering on GPUs, enabling robots to perceive their environment more efficiently during training. The framework also includes pre-implemented code for numerous RL locomotion techniques, such as Teacher-Student learning and Adversarial Motion Priors (AMP).

Who it’s for

Robotics researchers and engineers focusing on legged locomotion and reinforcement learning who need a flexible, simulator-agnostic environment for training and testing robot policies.

Highlights

• Multi-Simulator Support: Seamlessly switch between IsaacGym, Genesis, and IsaacSim depending on whether you prioritize speed or realism.
• Comprehensive Method Library: Includes implementations of various published RL papers, including DeepMimic, DreamWaQ, and Walk These Ways.
• GPU-Accelerated Perception: Integrated Warp-based depth image rendering for faster visual input processing.
• Broad Robot Support: Compatible with various robots including Unitree Go2, TRON1, and Booster K1.