A team of researchers has unveiled Genesis, an open source computer simulation system that accelerates robot training by up to 430,000 times compared to real-world testing. The system allows robots to practice tasks in simulated reality, reducing the need for expensive physical testing and enabling faster learning.
Genesis uses AI agents to generate 3D physics simulations from text prompts, creating virtual environments that can be tailored to specific tasks. This enables researchers to train neural networks for piloting robots in just hours of real computer time, equivalent to decades of training in the real world.
The platform’s developers say it has the potential to revolutionize robotics by providing a more efficient and effective way to test and train robots in virtual environments. By processing physics calculations up to 80 times faster than existing simulators, Genesis can handle complex simulations that were previously unfeasible.
Additionally, Genesis can generate “4D dynamic worlds” – complete virtual environments generated from text descriptions – allowing researchers to simulate a wide range of scenarios and test the limits of their robots. This breakthrough technology has significant implications for the development of real-world robots that can perform complex tasks with accuracy and precision.
Source: https://arstechnica.com/information-technology/2024/12/new-physics-sim-trains-robots-430000-times-faster-than-reality