The virtual worlds where robots are trained

Freddo, a robot developed by Cambridge‑based start‑up Vsim, demonstrated the ability to walk across an office, recognize a plastic bottle offered by a staff member and grasp it after only a few minutes of virtual training. The rapid acquisition of these skills, which the company claims rivals can take days to achieve, was made possible by Vsim’s proprietary simulation environment that runs directly on the robot’s hardware, allowing it to evaluate tens of thousands of possible future scenarios in real time. Founders Michelle Lu and Kier Storey, former contributors to Nvidia’s Isaac Sim, built the system from the ground up to exploit modern graphics processing units, achieving a “super high‑performance” simulator within eighteen months that can predict about a second ahead for 20,000 different combinations of events, a capability they argue is essential for navigating unstructured settings such as homes and workplaces.

The core advantage of Vsim’s approach lies in its speed and efficiency: by redesigning algorithms originally conceived in the 1970s and 1980s to run natively on GPUs, the simulator can generate millions of training iterations in the time traditional systems need for days. This acceleration enables the robot to develop a “policy” – an optimal set of actions – within a virtual world before uploading it to the physical unit. The company’s ten‑engineer team highlights that the simulator’s lightweight footprint lets Freddo run simulations on‑board, a contrast to industry heavyweight Nvidia, which supplies large‑scale AI chips and a suite of robotics software but still provides only a rudimentary grasp of real‑world physics. Nvidia’s own efforts, such as the Cosmos world model and AI agents that automate environment creation, aim to address similar challenges, yet both firms acknowledge that fine dexterity and long‑horizon tasks—like picking up a bottle, filling it, and pouring—remain difficult for robots.

Experts see Vsim’s fast‑simulation model as a promising complement to other training methods, such as learning from human demonstrations or using open‑source platforms like MuJoCo, which University of Cambridge associate professor Rika Antonova praises for its accessibility. Antonova notes that the ability to run hundreds of millions of samples quickly could bridge the gap between simple manipulation and more complex, sequential tasks. As Vsim continues to refine its technology, the implications extend beyond a single demo: faster, on‑board simulation could enable a new generation of robots capable of adapting instantly to dynamic, real‑world environments, potentially reshaping how automation is deployed in homes, offices, and industrial settings.

⚡ Effects Interpreter