Stories

Thu, 17 Sep 2026

Thu, 17 Sep 2026 The virtual worlds where robots are trained

Training systems that allow robots to negotiate the real world are getting more sophisticated.
The founders of Vsim, Michelle Lu and Kier Storey, attribute their success to the use of powerful computer chips called graphics processing units (GPUs) to accelerate the training process. Their system can run tens of thousands of simulations on Freddo's hardware while it is moving around, allowing it to adapt quickly to changing situations. This approach is more efficient than traditional methods used by tech giants like Nvidia, which requires manual labor to build virtual environments and scan them in. Vsim's software can create complex simulations much faster and with greater accuracy, enabling robots to learn complex behavior more quickly. Rika Antonova, an associate professor at the University of Cambridge, praises Vsim's approach as "promising" but notes that simulated environments are still rough approximations of the real world, limiting what can be trained. However, Lu says their system has reduced approximation and can train models that genuinely work in reality as well as they do in simulations. A second robot, Nacho, is expected to join Freddo soon to further develop the technology and ensure it can run on different machines.