rl_games

A reinforcement learning training library optimized for massively parallel GPU simulation, common in the Isaac stack.

rl_games is an open-source reinforcement learning library by Denys Makoviichuk and collaborators, implementing mainly PPO (Proximal Policy Optimization) and SAC. Its defining feature is that training data can stay entirely on the GPU throughout, which suits simulators like Isaac Gym that run thousands of parallel environments at once. NVIDIA's IsaacGymEnvs uses it as the default training backend, and Isaac Lab lists it as one of its supported training libraries; the open-source code behind many dexterous-hand and legged-locomotion papers is built on it. Compared with rsl_rl, it has more features but more configuration surface, and hyperparameters are usually written in YAML files.

ExampleRun rl_games' training script inside Isaac Lab with a task name specified, and it will train a robot hand to reorient a cube with PPO across thousands of parallel environments.

Related

Proximal Policy Optimization · Massively Parallel Reinforcement Learning · NVIDIA Isaac Lab · Isaac Gym · rsl_rl · skrl

Sources

rl_games on GitHub