Vsim, a Cambridge-based robotics startup founded by Michelle Lu and Kier Storey, has created a software system designed to accelerate robot training through virtual simulation. The technology enables tasks to be learned and optimized in computer simulations before being deployed on physical robots. Freddo, a demonstration robot, was trained to walk, recognize, and grasp a bottle in just minutes using this approach, compared to days required by competing systems.
The founders previously worked on Nvidia’s Isaac Sim platform before establishing Vsim in 2022. By building their system from the ground up, they optimized the underlying algorithms to work efficiently with graphics processing units (GPUs) used in artificial intelligence applications. Traditional robotics simulation algorithms dating to the 1970s and 1980s were not well-suited to modern GPU architecture. Within eighteen months, Lu and Storey developed a fully functional, high-performance simulator capable of running on Freddo’s onboard hardware.
The efficiency of Vsim’s system allows robots to execute tens of thousands of simulations while moving, enabling them to anticipate potential scenarios one second into the future. This capability is critical for robots operating in unstructured environments like homes, where unpredictable elements such as humans, animals, or other robots may require rapid strategy adjustments. The ten-engineer startup operates at a vastly smaller scale than Nvidia, which dominates AI chip markets and employs hundreds of engineers in its robotics division. Nvidia’s robotics software suite includes simulation systems and a world model called Cosmos that provides robots with physics understanding.
Other training approaches exist alongside simulation. Researchers at Cambridge, including associate professor Rika Antonova, utilize MuJoCo, an open-source training system owned by Google’s DeepMind since 2021. While simulation-based training shows promise, challenges remain in modeling complex real-world phenomena such as deformable objects and cutting operations. Both Nvidia and Vsim are actively addressing these limitations. A second robot named Nacho will assist in further developing technology that bridges the gap between simulated and real-world performance.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI