Brax: A New Differentiable Physics Engine

This time your hosts Sanyam Bhutani & Cristian Garcia are joined by Daniel Freeman, a senior Software Engineer at Google Research. In this talk, we introduce Brax [arxiv] [GitHub] [blogpost], a new differentiable physics engine that we developed in JAX. We’ll provide some motivating context for the engine, an interactive demo showcasing how Brax can greatly accelerate training speed on accelerators, design details describing how Brax works at a high level, an example of how Brax can greatly improve sample efficiency by leveraging differentiability, and finally some simple comparisons between Brax and several other popular engines. Get started with W&B: ​ Follow us: Twitter: Linkedin: ​