leggedrobotics/rsl_rl

What it solves

RSL-RL provides a lightweight, GPU-accelerated learning library specifically designed for robotics research. It removes the complexity and overhead found in larger libraries, allowing researchers to prototype and test new robot learning ideas quickly without needing to modify extensive codebases.

How it works

The library is designed with a minimal and readable codebase that offers clear extension points for rapid prototyping. It implements common robot learning algorithms, such as Proximal Policy Optimization (PPO) and Student-Teacher Distillation, and supports native multi-GPU training to achieve high-throughput training speeds.

Who it’s for

It is primarily intended for robotics researchers who need a flexible, compact tool for implementing and testing new learning algorithms in simulated environments.

Highlights

• GPU Acceleration: Native multi-GPU support for high-throughput training.
• Robotics-First Methods: Includes built-in support for PPO and Student-Teacher Distillation.
• Minimal Design: A compact, readable codebase that simplifies the extension and extension of algorithms.
• Broad Integration: Used by major robot learning environments including Isaac Lab, Legged Gym, mjlab, and MuJoCo Playground.