World Labs Trains Robots in Simulation for Real-World Tasks
TL;DR. World Labs released a simulation engine that trains robot controllers in virtual environments, then transfers the models to real hardware. - The system generates thousands of task variations from a single real-world example to improve generalization. - Trained models operated for an hour on various robot platforms without human intervention. - This approach aims to reduce the cost and difficulty of real-world robot data collection.
- World Labs released a simulation engine (R2S2R) for training robot control systems.
- The engine converts real-world robot tasks into interactive virtual environments.
- It generates thousands of controlled variations (lighting, object position, friction) from one real task.
- Control models are trained entirely in simulation and then deployed to real robots.
- Test models ran for one hour on five different robot platforms without human input.
- The technology aims to overcome the data bottleneck in robot deployment by using simulation.