google/brax

Health v2 · maintenance only

Adoption not part of the score

3223 stars · 347 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Brax is a fast, fully differentiable rigidbody physics engine written in JAX, designed for massively parallel simulation on accelerator hardware like TPUs and GPUs. It also includes a suite of reinforcement learning training algorithms (PPO, SAC, ARS, ES, APG), which is now the actively maintained part of the project.

Use cases

• train reinforcement learning agents for robotics in minutes
• run massively parallel physics simulation on GPU or TPU
• differentiable physics simulation for research
• simulate humanoid and quadruped environments at millions of steps per second
• train RL policies with PPO or SAC on simulated environments
• use analytic policy gradients with a differentiable simulator

When to choose

• you need differentiable or massively parallel physics simulation in JAX
• you want to train RL agents quickly on accelerator hardware
• you are doing robotics or RL research with GPU/TPU access

When to avoid

• you need a general-purpose physics engine for production robotics simulation - use MuJoCo MJX or MuJoCo Warp instead
• you need the brax/envs pipeline - it is no longer maintained
• you have no accelerator hardware, since Brax is designed for TPU/GPU

Facets

library · maturity maintenance

simulation machine-learning reinforcement-learning gpu-computing robotics reinforcement-learning simulation machine-learning python cross-platform jax differentiable-physics physics-engine ppo sac mujoco tpu robotics-simulation gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="google/brax")