rsl_rl

ETH Zurich's open-source reinforcement learning library built for GPU-parallel simulation.

rsl_rl is a lightweight, open-source PyTorch reinforcement learning library from the Robotic Systems Lab (RSL) at ETH Zurich, originally built to train quadruped locomotion in Isaac Gym alongside legged_gym. At its core is a PPO (Proximal Policy Optimization) implementation optimized for thousands of parallel environments, with all sampling and update data kept on the GPU to avoid costly CPU round-trips. It's now one of the RL libraries Isaac Lab officially supports out of the box, and it's also used by legged- and humanoid-locomotion projects such as unitree_rl_gym; later versions have added recurrent-network policies, symmetry augmentation, and teacher-student distillation. Reproducing a legged-locomotion paper today can hardly avoid it.

ExampleUse rsl_rl's training script in Isaac Lab to train a velocity-tracking policy for the Unitree Go2, sampling thousands of environments in parallel on the GPU before updating with PPO.

Also called

rsl-rl, RSL RL

Related

legged_gym · NVIDIA Isaac Lab · Proximal Policy Optimization · Massively Parallel Reinforcement Learning · ETH Zurich Robotic Systems Lab · rl_games

Sources

leggedrobotics/rsl_rl (GitHub)