British Startup Trains Robot Freddo to Learn Skills in Minutes

A British technology startup says its robot training system can teach robots new physical skills in minutes, potentially accelerating the development of useful machines for homes and workplaces.

At Cambridge-based Vsim, the humanoid robot Freddo recently demonstrated the technology by walking across an office, recognising a plastic bottle and grasping it from a staff member.

The task may appear simple, but Vsim says the speed of the training process sets its system apart. Developers trained Freddo to perform the sequence within minutes, while competing systems can take days to develop similar skills.

Vsim Uses Virtual Training

Founders Michelle Lu and Kier Storey hope their software will eventually help robots navigate unpredictable environments and perform useful everyday tasks.

For now, however, significant technical challenges remain.

“It’s a weird situation with robotics because actually the stuff that we find as humans to be incredibly difficult, like gymnastics, you can get robots to do reasonably well,” Storey said. “The stuff that humans are really good at, like fine dexterity, is really hard in robots.”

OpenAI to Regularly Report AI Misbehavior in New Transparency Push

Vsim trains Freddo inside a virtual environment. A computer simulation can repeat the same task millions of times while searching for the most effective way to complete it.

Once the system identifies the optimum solution, developers can transfer that policy to the physical robot.

GPU Technology Speeds Up Robot Training

Virtual simulation has become a common tool for robotics developers. Nvidia, one of the world’s leading AI chip companies, also offers its Isaac Sim platform for robot training.

Lu and Storey previously worked on an early version of Nvidia’s system. They launched Vsim in 2022 to develop their own simulation environment and robotics tools.

Starting from scratch allowed the founders to optimise their software for graphics processing units, or GPUs. Those chips power many modern AI systems.

Storey said many algorithms used in robotic simulation date back decades and do not fully exploit modern GPUs.

Within months, the team found that its approach could run simulations much faster than systems they had previously encountered.

Freddo Can Run Simulations on the Move

Vsim says its software can run directly on Freddo’s onboard hardware. That allows the robot to perform tens of thousands of simulations while moving through its surroundings.

Storey said Freddo can examine around 20,000 possible combinations of future events while looking roughly one second ahead.

Such rapid decision-making could prove important in homes and other unpredictable environments.

People, animals or other robots could suddenly change their behaviour. A useful household robot would therefore need to adjust its actions quickly while maintaining safety.

Nvidia Develops AI Tools for Robotics

Vsim has around 10 engineers working on its technology. Nvidia, by contrast, employs hundreds of engineers across its robotics software operations.

The chipmaker does not manufacture robots. Instead, it provides software that organisations can use to train and control robotic systems.

Its tools include virtual simulation platforms and Cosmos, a so-called world model designed to help robots understand physical environments and predict how those environments may change.

Nvidia robotics executive Spencer Huang said simple manipulation tasks remain manageable. More complicated assignments become difficult when they involve several steps.

For example, asking a robot to pick up a bottle is relatively straightforward. Telling it to fill the bottle, carry it somewhere and pour the contents requires much longer-term planning.

Nvidia has also begun using AI agents to create virtual environments and test robotic training results. Huang said the approach could reduce the manual work involved in building simulation environments.

Researchers Explore New Ways to Train Robots

Simulation is not the only way to teach robots. Researchers can also use human demonstrations and video footage to train machines.

Rika Antonova, an associate professor at the University of Cambridge, has worked in robotics since 2015. Her research focuses on software and hardware that help robots learn complex behaviours.

Antonova uses MuJoCo, an open-source robotics simulation platform owned by Google’s DeepMind.

She said Vsim’s fast-simulation approach shows promise because it could allow robots to process hundreds of millions of simulated examples while deciding how to move.

However, simulation still cannot perfectly reproduce the physical world.

Highly deformable materials and tasks such as cutting remain difficult to model accurately. These limitations can affect how well skills learned in a virtual environment transfer to real robots.

Vsim Plans Second Robot

Lu said Vsim is working to reduce those gaps by improving the accuracy of its simulations.

The company plans to introduce a second robot, named Nacho, to support development. Vsim expects the additional machine to accelerate testing and help ensure its software works across different robotic platforms.

For Freddo, the new robot will also provide something less technical: company.