British start-up Vsim, based in Cambridge, has developed a high-performance virtual simulation system designed to train robots quickly. Founders Michelle Lu and Kier Storey claim their software can develop robot skills in minutes, a process they say could take days with rival systems.
The system trains robots in a virtual environment, allowing tasks to be performed millions of times in simulation to find an optimum solution. This solution, known as a policy, can then be uploaded to physical hardware, such as the robot Freddo, which was trained to walk and grasp a bottle.
Vsim's software was optimised to exploit graphics processing units (GPUs), powerful computer chips used in AI. This efficiency allows the software to run on the robot's hardware, enabling it to perform tens of thousands of simulations while moving and anticipate future events.
The company aims for its software to control robots performing useful tasks in homes and workplaces, particularly in unstructured environments where robots need to adapt quickly to unexpected events like the presence of humans or animals. Vsim states its system uses accurate simulations to train models that work effectively in both virtual and real-world settings.