Genesis, a just-released open source physics simulator, pushes the boundaries of virtual environments for robot training and embodied AI development, with performance far superior to existing solutions.
A group of academic and industrial researchers unveiled Genesis, a
simulator of physical environments with unprecedented performance. This open source platform is designed for applications in the fields of robotics and embodied artificial intelligence.
43 million frames per second
According to the official project documentation, Genesis uses an optimized physics engine that harnesses the power of GPU parallel computing. This technology makes it possible to achieve an unprecedented simulation speed, up to 43 million frames per second (FPS) in certain manipulation scenes. Robots can thus train in a virtual environment up to 430,000 times faster than in the real world. Running entirely in Python, Genesis far exceeds the performance of existing solutions such as Nvidia’s Isaac Gym, with speeds up to 80 times faster, the project’s creators claim. Using graphics cards similar to those used for video games, the system can run up to 100,000 copies of a simulation environment simultaneously.
Unified simulation framework
The team behind Genesis explains that the platform can generate “dynamic 4D worlds” (a simulation of a 3D world with the temporal component). The stated goal is to build a universal engine capable of automatically generating complex physical worlds and scenarios. Including interactive environments, camera movements, robotic tasks but also robot control policies and fully interactive 3D scenes. Zhou Xian, a researcher at Carnegie Mellon University and a key contributor to the project, points out on X that Genesis is the result of a 24-month collaboration. “Genesis implements a unified simulation framework, integrating a wide range of cutting-edge physics solvers, to simulate the physical world with maximum realism,” he adds.
Virtual environments created via prompts
A functionality is also based on generative AI, allowing the automatic creation of complete virtual environments from simple textual descriptions (prompts). The project’s official website features several examples of 3D environments accurately rendered using these types of descriptions. “A Franka mobile arm rearranges the books on a table by pushing the brown and white books to align them with the red book,” illustrates a robotic training use case. Genesis can also simulate soft or hybrid robots (combining soft skin and a rigid skeleton). Example prompt: “A flexible gripper navigates a box and grabs a pen cap.”
Note also that Genesis benefits from cross-platform compatibility, with native support for Nvidia, AMD, Apple and Intel GPUs and CPUs, on Windows, macOS and Linux.
Tech
