Virtual simulations speed up robot training, start-up says

Cambridge start-up Vsim is developing software that trains robots in virtual environments before transferring learned skills to physical machines. Its founders say the system can run thousands of simulations quickly, although researchers caution that simulated worlds still struggle to reproduce complex real-world tasks.

Freddo, a robot at Cambridge-based start-up Vsim, can walk across an office, identify a plastic bottle and take it from a person. The company’s founders say those abilities were developed and uploaded in minutes, a process they describe as faster than rival systems that can take days.

Vsim founders Michelle Lu and Kier Storey are building software intended to help robots navigate homes and workplaces. Their approach trains robots in computer-generated environments, where a task can be repeated millions of times to find an effective solution, or “policy”, before it is transferred to a physical machine.

The founders say Vsim’s software was designed to make better use of graphics processing units, or GPUs, which are widely used for artificial-intelligence workloads. Storey said many algorithms used in robotic simulations date from the 1970s and 1980s and are not well suited to GPUs. Lu said the company had developed a functional high-performance simulator within 18 months.

The software can also run on hardware carried by Freddo, allowing the robot to evaluate tens of thousands of possible scenarios while moving. Storey said it can look roughly a second ahead and assess 20,000 combinations of events. Lu said this could help robots react when people, animals or other machines behave unexpectedly in an unstructured environment.

Nvidia, which has a much larger robotics software operation, also provides simulation tools and a system called Cosmos that is intended to help robots understand real-world physics and changing surroundings. Spencer Huang, Nvidia’s director of product for robotics, said current systems remain limited, particularly for extended tasks involving several steps. Nvidia has begun using AI agents to help create virtual environments and check whether training results work as intended.

University of Cambridge associate professor Rika Antonova said fast simulation was promising because it could allow robots to test hundreds of millions of examples in seconds. But she also warned that simulations remain approximations, with highly deformable objects and cutting among the difficult phenomena to model. Vsim says it is working to reduce those gaps, and plans to use a second robot, Nacho, to help test software across different machines.