Freddo, a robot, walks across an office and takes a plastic bottle offered by a staff member. While this may not seem groundbreaking—especially after a robot recently beat the 100m sprint record set by Usain Bolt—it highlights how quickly Freddo was trained to walk, recognize the bottle, and grasp it. The process took just a few minutes, a feat that rival systems might require days to achieve. This development is happening at Vsim, a British start-up based in Cambridge, founded by Michelle Lu and Kier Storey. Their goal is to create software that can eventually control robots capable of performing practical tasks in both homes and workplaces. Despite their progress, Storey notes that robotics still face challenges. "As humans, we find tasks like gymnastics difficult, but robots can handle them reasonably well. However, tasks that we humans excel at—like fine dexterity—are extremely hard for robots," he explains. To overcome these challenges, Freddo’s skills were developed in a virtual environment where tasks can be simulated millions of times. Once an optimal solution, or "policy," is found, it can be uploaded to the robot. This method of virtual training is common in the field, and both Lu and Storey previously worked on an early version of Nvidia’s Isaac Sim, a similar system. What sets Vsim apart is how they optimized their software to work efficiently with powerful computer chips used in AI, known as graphics processing units or GPUs. Storey points out that many of the algorithms used in robotic simulations are decades old and not well-suited for modern GPUs. Within a few months, Vsim’s system showed significant improvements in speed. "Eighteen months in, we actually have a completely functional, super high-performance simulator," Lu says. The efficiency of Vsim's software allows Freddo to run tens of thousands of simulations while moving around. This enables the robot to predict and adapt to various scenarios in real time, a crucial ability for navigating unpredictable environments like a home. Meanwhile, Nvidia, a major player in the AI chip and robotics software market, also uses simulation training but faces its own challenges. While Nvidia's software provides a basic understanding of the real world, more complex tasks—like pouring water into a bottle—remain difficult to simulate accurately. However, Nvidia is making progress by using AI agents to help build and validate virtual environments for training robots. Simulation is one of several methods for training robots. Researchers like Rika Antonova, an associate professor at the University of Cambridge, also explore other techniques, such as teaching robots by watching human demonstrations. Antonova notes that Vsim’s fast simulation approach is promising, allowing for real-time adjustments. However, she cautions that simulated environments are still approximations of the real world, making certain tasks—like handling highly flexible objects—challenging. Both Nvidia and Vsim are working to bridge this gap. Lu mentions that a second robot, named Nacho, will soon join Freddo in development, helping to refine their technology and ensure it works across different machines.