Robot Training Systems Evolve for Real-World Tasks

Robots are becoming increasingly adept at performing complex tasks, a development fueled by advancements in training systems that utilize sophisticated virtual environments. British startup Vsim is among those pushing the boundaries, developing software that can train robots in simulated worlds in mere minutes, a process that previously could take days.

Vsim's approach focuses on optimizing training algorithms for powerful graphics processing units (GPUs), enabling their software to run simulations at remarkable speeds. "Eighteen months in and we actually have a completely functional, super high-performance simulator," said co-founder Michelle Lu. This speed allows robots to run tens of thousands of simulations while performing tasks, enabling them to anticipate potential future scenarios and adapt quickly to unexpected events.

This rapid simulation is crucial for robots operating in unpredictable environments, such as homes. "Things outside of the robot's control, like humans, animals or even other robots, could do things that require a change of strategy," explained co-founder Kier Storey. "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."

Tech giant Nvidia is also heavily invested in robot training, offering systems like Isaac Sim that create virtual environments for training. Nvidia's Spencer Huang, director of product for robotics, highlighted the company's efforts in using AI agents to build and validate virtual environments. "We're just throwing agents at it... it's basically given us a huge workforce," Huang said, referring to the use of AI in creating training scenarios.

However, replicating the nuances of the real world in simulation remains a significant challenge. "The stuff that humans are really good at, like fine dexterity, is really hard in robots," Storey noted. Huang echoed this, stating that while simple manipulation like grasping a bottle is manageable, long-horizon tasks, such as filling a bottle and pouring its contents, are more difficult. Specific real-world elements like highly deformable objects and cutting are also challenging to model accurately in simulations, according to Rika Antonova, an associate professor at the University of Cambridge's Department of Computer Science and Technology.

Antonova's research utilizes MuJoCo, an open-source training system owned by Google's DeepMind. She acknowledged the promise of Vsim's fast simulation approach, stating, "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."

Companies like Vsim and Nvidia are working to reduce the gap between simulated training and real-world performance. Vsim aims to use accurate simulations to train models that perform as well in reality as they do in virtual environments. The addition of a second robot, Nacho, is expected to accelerate Vsim's development process and ensure their software's compatibility across different machines.