Virtual worlds accelerate robot training

BBC reports that Vsim, a British start-up based in Cambridge, is using virtual environments to train robots more quickly. Its robot Freddo learned to walk, recognise a plastic bottle and grasp it in a few minutes. The company's founders, Michelle Lu and Kier Storey, say rival systems could take days.

In simulation, a task can be repeated millions of times until an effective solution, known as a policy, is found. That policy can then be uploaded to a robot. Vsim was founded in 2022 after Lu and Storey worked on an early version of Nvidia's Isaac Sim. They designed their system to use graphics processing units, or GPUs.

Lu says that, after eighteen months, Vsim had a functional, high-performance simulator. The software runs on Freddo, allowing it to perform tens of thousands of simulations while moving. Storey says it can consider about one second ahead across 20,000 possible combinations, which could help robots adapt in homes where people, animals and other robots create unexpected events.

Nvidia also develops simulation and world-model software, including Cosmos. Rika Antonova of the University of Cambridge says Vsim's speed is promising, but simulations remain approximations. She notes that highly deformable objects and cutting are difficult to model. Vsim also works to reduce that gap between simulation and reality.