Genesis Open-Source AI Physics Engine Introduced, Can Generate 4D Dynamic Worlds to Train Robots

Genesis, a generative artificial intelligence (AI) physics model that can simulate four-dimensional (4D) worlds, was unveiled on Thursday. This is a unique AI model, as it combines several different capabilities to simulate scenarios for general-purpose robotics and physical AI applications. The researchers behind the project claimed that Genesis excels at simulation speed and is up to 80 times faster than typical GPU-accelerated systems. Notably, the open-source AI system is available via the Python Package Index (PyPI) repository. But those installing will need to install PyTorch as well.

Genesis AI Physics Model Can Simulate Dynamic Worlds for Robotics Training

In a post on X (formerly known as Twitter), Zhou Xian, the lead researcher of the project, announced Genesis and highlighted that it was built after a two-year-long large-scale research collaboration including more than 20 research labs. It integrates multiple physics solvers and their coupling into a unified framework.

Built entirely on Python, it features a generative agent framework and is powered by a universal physics engine. Currently, the group has only open-sourced the underlying physics engine and the simulation platform. It stated that the generative framework will be released in the future.

The promise of this AI system is big, as per the claims made by the researchers. It is said to be between 10 and 80 times faster than systems like Isaac Gym and MJX, which rely on GPU acceleration to create simulations. Further, in specific scenarios, the engine is claimed to deliver 4,30,000 faster simulation speed than real-time. The lead researchers added that it can train a robotic locomotion policy on a single Nvidia RTX4090 GPU in 26 seconds.

Coming to the key features of the Genesis AI physics engine, it is fully integrated with Python as both the frontend and the backend of the engine were natively developed in it. The system is also available via an application programming interface (API). Despite faster simulation speeds, it is said to maintain simulation accuracy and fidelity. Its unified framework also enables multiple physics solvers to create a wide range of physical phenomena and materials. The physics engine also offers ray-tracing rendering.