“We already know that certain organs accelerate learning,” says Bongard. “This work shows that AI can search for such bodies.” Bongard’s lab has developed robotic bodies that are tailored to specific tasks, such as applying callus-like coatings to their feet to reduce wear and tear. Gupta and his colleagues are taking this idea further, says Bongard. “They show that the right body can also accelerate changes in the robot’s brain.”
Ultimately, this technique could completely change the way we think about building physical robots, Gupta says. Instead of starting with a fixed body configuration and then training a robot to perform a specific task, you can use DERL to allow the optimal body plan for that task to evolve and then build it.
Gupta animals are part of a massive shift in how researchers think about AI. Instead of teaching AIs specific tasks like playing go or analyzing medical scans, researchers are starting to throw bots into virtual sandboxes like POET, OpenAI’s virtual hide-and-seek arena, and XLand DeepMind’s virtual playground, and get them. to learn how to solve multiple problems in an ever-changing training dojo with an unlimited duration. Instead of tackling a single task, AIs trained in this way learn general skills.
For Gupta, free-form research will be key to the next generation of AI. “We need truly open environments to build intelligent agents,” he says.