Cambridge Start-up Vsim Trains Robots in Minutes, Not Days

A British start-up says it taught a robot to walk, recognise a plastic bottle and grasp it in just a few minutes — a task rival systems could take days to complete.

The robot, called Freddo, walks across the office of Vsim, a Cambridge-based firm founded in 2022 by Michelle Lu and Kier Storey. The founders, who both worked on an early version of Nvidia's Isaac Sim robotics platform, left to build their own training environment from scratch. Their goal: robots that can navigate and perform useful tasks in homes and workplaces.

Freddo's skills were honed in a virtual environment, where a task can be repeated in computer simulation millions of times. Once the optimum solution — known as a policy — is found, it uploads directly to the hardware.

Because Lu and Storey started with a clean slate, they could optimise the software for the graphics processing units (GPUs) that power modern AI. Storey argues the industry's legacy tools were holding it back.

"The underlying algorithms that we were using for most of these robotic simulations they hark back to the 1970s and 1980s, but those algorithms are not really brilliant fits for GPUs," Storey says.

The bet paid off fast. "Eighteen months in and we actually have a completely functional, super high-performance simulator," Lu says.

The software is efficient enough to run on the hardware Freddo carries. The robot can run tens of thousands of simulations while moving around.

"It can look about a second, or so, ahead into the future for 20,000 different kind of combinations of things that might happen," Storey explains.

That matters in unstructured settings like a family kitchen. "Things outside of the robot's control, like humans, animals or even other robots, could do things that require a change of strategy. These unexpected events could happen very quickly and the robot needs to be able to quickly adapt to ensure its actions remain safe and on-mission," Lu says.

Vsim fields 10 engineers. Its competitor Nvidia sits at the other end of the industry: the chip giant dominates the AI semiconductor market and runs a robotics software division with hundreds of engineers. Nvidia does not build robots; it sells a suite of software for training and controlling them, including simulation systems and a "world model" called Cosmos, which gives robots an understanding of real-world physics.

Even with Nvidia's compute resources, the software delivers only a rudimentary grasp of reality.

"Manipulation, where I just grab a bottle, that's not too hard. The problem is when you start doing long-horizon tasks, where I say: 'I want you to take the bottle and I want you to fill it up and I want you to go pour'," says Spencer Huang, director of product for robotics at Nvidia.

This year Nvidia started deploying AI agents to build virtual training environments and validate whether trained solutions actually work. "When we talk about creating the [virtual] world and actually scanning it in — a lot of that is actually manual labour. We're just throwing agents at it... it's basically given us a huge workforce," Huang says.

Simulation is not the only training method. Robots can also learn from human or video demonstrations. Rika Antonova, an associate professor in Cambridge University's Department of Computer Science and Technology who has worked in robotics since 2015, uses MuJoCo, the open-source simulator owned by Google's DeepMind since 2021.

"It is very, very user-friendly. So for research groups or for small start-ups, that's useful," she says. She calls Vsim's fast-simulation approach promising: "If you have a very, very fast simulator, then you can simulate hundreds of millions of samples in that few seconds that your robot is thinking about how to adjust its motion, and then you can change the motion almost in real time."

The constraint remains fidelity. Simulated environments are rough approximations of reality. "There are certain things that are hard to model in simulation, like highly deformable objects and cutting," Antonova says.

That is exactly the problem Lu says Vsim is attacking, with a system that has "reduced approximation, using accurate simulations to train models that genuinely work in reality as well as they do in simulations."

A second robot, Nacho, joins the lab soon. Lu says it will speed development and prove the software runs on different machines — and give Freddo some company.