Genesis is an open source generative physics engine that can train robots in a simulated world 430,000 times faster than the real world

A research team at Carnegie Mellon University has announced Genesis , a physics simulation platform that can run simulations 430,000 times faster than reality. Genesis has a lightweight Python-based design, fast physics calculations, and world generation capabilities using natural language, and is available as open source.

Genesis
https://genesis-embodied-ai.github.io/

GitHub - Genesis-Embodied-AI/Genesis: A generative world for general-purpose robotics & embodied AI learning.
https://github.com/Genesis-Embodied-AI/Genesis

New physics sim trains robots 430,000 times faster than reality - Ars Technica
https://arstechnica.com/information-technology/2024/12/new-physics-sim-trains-robots-430000-times-faster-than-reality/

The simulation generated by Genesis can be seen in the video below.

'Genesis', an open source generation physics engine that can train robots in a simulated world 430,000 times faster than the real world, looks like this - YouTube


Genesis has been designed from the ground up as a general-purpose physics engine, with optimized collision detection, auto-pause, contacting islands and other features for efficient GPU parallel computing.

Genesis also uses Python for both the interface and the core physics engine, allowing for fast robot training simulations to be run with simple Python commands on commodity hardware. It also leverages a visual language model (VLM) to generate complete virtual environments from text prompts, creating a 4D dynamic world with camera movements and object behavior that follow the laws of physics.

Additionally, the Genesis rendering system can generate physically accurate ray-traced videos and data for use as training data for robots, as well as character motion, interactive 3D scenes, facial animations, and emotional expressions.

Below is how Genesis generates a 3D space for robot training.

'Genesis' can generate interactive 3D scenes for robot training - YouTube


And the following movie shows the facial animation reproduced by Genesis.

Reproducing facial animation with 'Genesis' - YouTube


For example, the simulation scene below using a robot arm achieved an astounding processing speed of 43 million frames per second, which is 430,000 times faster than reality. The neural network controlling the robot can virtually gain decades of learning experience in just a few hours of real computer time.

Simulating a robot arm with the open source physics engine 'Genesis' - YouTube


Genesis can also generate 4D dynamic worlds from natural language instructions, including physically accurate camera movements, object behavior, etc. Genesis' physics engine and simulation platform are available as open source, and the generation framework will be released in stages in the future.

Many renowned research institutes and companies, including Carnegie Mellon University, Massachusetts Institute of Technology, Stanford University, and NVIDIA, are participating in this project, which has the potential to greatly accelerate research and development in robotics. The development team is offering Genesis free of charge, based on the idea that robotics is a grand project that should be tackled by the entire human race.