A British start-up based in Cambridge has developed a robotics training system that can master new physical tasks in minutes rather than days, according to its founders. Vsim's approach uses high-speed virtual simulations to train robots like Freddo, a humanoid unit that recently learned to walk, recognize a plastic bottle and grasp it after only a few minutes of digital practice.

The company's founders, Michelle Lu and Kier Storey, describe their method as exploiting modern graphics processing units to run millions of simulated trials within seconds of real-world robot movement. "Our system can look ahead about a second into 20,000 possible future scenarios while the robot operates in unstructured environments," Storey explained. This capability could prove vital for robots navigating homes or workplaces where unexpected human or environmental interactions require rapid adaptation.

Vsim's technology builds on decades-old simulation algorithms but re-engineered them to leverage contemporary GPU architectures. The founders noted that traditional robotics simulation software often fails to align with GPU capabilities, creating performance bottlenecks. Their system now processes tens of thousands of simulated motion combinations per second, enabling real-time strategy adjustments during physical tasks.

The company operates with a team of 10 engineers and emerged from research conducted in 2022. Their work contrasts with approaches from major technology firms like Nvidia, which dominates AI chip markets and maintains extensive robotics software tools including Isaac Sim and Cosmos world modeling platform. Nvidia's Spencer Huang acknowledged ongoing challenges with long-horizon tasks such as multi-step object manipulation, though he noted progress through AI agents that generate virtual training environments.

Academic perspectives highlight both promise and limitations. Rika Antonova, an associate professor at the University of Cambridge working with DeepMind's MuJoCo simulation platform, described Vsim's speed as "very promising" for enabling hundreds of millions of training samples per second. However, she noted that even advanced simulations struggle with modeling real-world complexities like deformable objects and precise cutting tasks.

Vsim plans to expand testing with a second robot named Nacho to evaluate cross-platform compatibility. Their development occurs amid broader industry efforts to improve robot training efficiency, as companies race to create systems capable of reliable operation in unpredictable physical environments.