The virtual worlds where robots are trained
BBC News
A Cambridge startup trained a robot to grab a bottle in minutes, not days. It’s a sign simulation is getting fast enough to steer real machines.
Based on reporting by BBC News — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Freddo the robot can walk over, spot a plastic bottle and grasp it after only a few minutes of training. That’s the eye-catching bit here. His developers say similar systems can take days to learn the same trick, which makes the speed jump stand out more than the task itself.
The company behind it is Vsim, a British start-up in Cambridge founded by Michelle Lu and Kier Storey in 2022. Their bet is that robots will get useful much faster if most of the learning happens inside a virtual world first, where the same task can be run millions of times before the result is sent to the real machine. Storey says that matters because robots are still awkward at the things people do instinctively, especially fine dexterity.
Vsim says it built its simulator from scratch so it could make better use of GPUs, the chips that power a lot of AI work. Storey says the older simulation algorithms go back to the 1970s and 1980s, which makes them a poor match for modern hardware. Lu says that after 18 months the team had a fully functional, high-performance simulator, and that it is efficient enough to run on Freddo itself while the robot is moving.
That gives the machine a chance to look roughly a second ahead across 20,000 possible outcomes, which sounds abstract until you picture a robot trying not to blunder into a chair, a person, or something else that moves. Lu says unstructured spaces like homes are exactly where robots need that kind of quick adjustment. And that is the real problem: not grabbing a bottle, but deciding what to do when the world changes under its feet.
Nvidia is pushing in the same direction with Isaac Sim and a wider robotics software stack, including a world model called Cosmos. Spencer Huang, who handles robotics product at Nvidia, says manipulation is manageable, but long tasks like taking a bottle, filling it and pouring it are much harder. Nvidia has also started using AI agents to help build and check virtual environments, because a lot of that work still takes manual labor.
Rika Antonova at the University of Cambridge says Vsim’s very fast simulation approach looks promising, especially for research groups and small start-ups. She also points out the hard limit: simulation still struggles with things like highly deformable objects and cutting. Vsim says a second robot, Nacho, is coming soon to help speed development and test whether the software works across different machines.
My take — AI-written commentary, not fact-checked reporting
This is the part of robotics that deserves the hype: not shiny demo bots, but faster ways to train them before they embarrass themselves in the real world. The industry has spent years acting as if more brute force would fix messy physical reality; it turns out better simulation may be the less glamorous, more useful answer. Tiny miracle, no cape required.
Read more about this at: BBC News