brax

GPU-accelerated NeuroEvolution of Augmenting Topologies (NEAT)

The Simple Simulator: Simulation Made Simple

MetaDE is a GPU-accelerated evolutionary framework that optimizes Differential Evolution (DE) strategies via meta-level evolution. Supporting both JAX and PyTorch, it dynamically adapts mutation and crossover strategies for efficient large-scale black-box optimization.

Walkthroughs for DSL, AirSim, the Vector Institute, and more

a modular reinforcement learning library with JAX agents

An implementation of Short Horizon Actor Critic writen in Jax. Core algorithm written in the style of Brax, with several bits taken from Xu's original paper.

GPU-accelerated Evolutionary Multiobjective Optimization Using Tensorized RVEA.

Brax Viewer is a real-time, interactive web viewer for monitoring reinforcement learning (RL) policies

Reinforcement learning in pure JAX.

💡 Grasp - Pick-and-place with a robotic hand 👨🏻‍💻

Simple implementations of Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP) in JAX

Relentlessly learning, persistently failing, but never surrendering.

[ICML 2024] Official environments and JAX-implementations for "Subequivariant Reinforcement Learning in 3D Multi-Entity Physical Environments"

Unitree Go2 terrain locomotion trained with MJX / MuJoCo Warp + Brax PPO on a single 8 GB GPU - 20.97M steps, 1536-768 MLP, stairs 0.03-0.05 m

[ICML 2026] Debate2Create: Robot Co-design via Multi-Agent LLM Debate

四足机器人 Go2 全栈控制：Convex MPC · RL（BC/PPO/域随机化）· 混合仲裁，同一套 72 试验推力协议下恢复率 94% —— 全程 CPU 训练，附与代码逐项一致的数学设计文档

APA-7 powered academic document editor online. Branded as "Acado".

A small humanoid robot you can talk to: walks with a reinforcement-learned policy, waves and dances, shows a face with lip sync and speaks - its brain is a local LLM you run yourself (Ollama or any OpenAI-compatible API). MuJoCo simulation, sim-to-real.

Procedural Environment Generation for Accelerated Multi-Agent Reinforcement Learning

Training a Unitree G1 Humanoid Robot to walk using PPO, JAX, and Brax. Achieved via End-to-End GPU Simulation (4096 parallel environments) and 200M timesteps of Curriculum Learning.