Cambridge-based start-up Vsim is developing high-performance simulation software designed to accelerate the training of robots for home and workplace tasks. By optimizing algorithms to leverage the power of graphics processing units (GPUs), the company has created a system capable of running tens of thousands of simulations in seconds. This allows a robot to anticipate potential environmental changes and adapt its strategy in real-time.

Co-founders Michelle Lu and Kier Storey, both former contributors to early versions of Nvidia’s Isaac Sim, established Vsim in 2022. The company’s current focus is on overcoming the difficulty robots face with fine dexterity and unstructured environments. According to Storey, while robots can perform complex physical feats like gymnastics, tasks requiring human-like precision remain a significant challenge. The Vsim team is currently working to ensure their simulations accurately reflect real-world physics, a goal they hope to further with the addition of a second robot, named Nacho, to their development team.

Advertisement

The broader robotics industry is also heavily invested in simulation. Nvidia, a major player in the field, utilizes its own suite of tools and a world model called Cosmos to help robots understand their surroundings. Spencer Huang, Nvidia’s director of product for robotics, noted that while simple tasks like grasping objects are becoming easier, executing "long-horizon" sequences—such as filling and pouring a bottle—remains a complex hurdle. To address this, Nvidia is increasingly employing AI agents to automate the creation of virtual training environments.

Academic researchers are also exploring various simulation methods. Rika Antonova, an associate professor at the University of Cambridge, utilizes the open-source MuJoCo system, which is owned by Google’s DeepMind. While she views Vsim’s high-speed simulation approach as promising for real-time motion adjustment, she notes that certain physical interactions, such as cutting or handling deformable objects, remain difficult to replicate accurately in a virtual space.

Source: BBC News