Overview
Brax is a fast and fully differentiable physics engine written in JAX, designed for research and development in robotics, human perception, materials science, and reinforcement learning. It supports efficient single-device simulation and massively parallel simulation on multiple devices, particularly excelling on TPUs, achieving millions of physics steps per second. Brax offers four physics pipelines: MuJoCo XLA (MJX), Generalized coordinates, Positional dynamics, and Spring-based simulation. It includes baseline learning algorithms like PPO, SAC, ARS, and analytic policy gradients. Brax facilitates experiments in transfer learning and bridging the simulation-to-real-world gap, offering an API for seamless pipeline swapping. While primarily an RL library, it encourages users to adopt MJX or MuJoCo Warp for pure physics simulations.