Consumer virtual reality has been a flop so far—but that doesn’t mean the technology is worthless.
Low-cost VR headsets could be the key to teaching industrial robots how to do more complex jobs in a shorter amount of time says Pieter Abbeel, president and co-founder of the startup Embodied Intelligence. Abbeel has the track record to back up those claims—the roboticist and UC Berkeley professor left the Elon Musk-backed research company OpenAI to start the venture, along with colleagues from Berkeley and the Musk lab.
Through the AI-based software the company is developing, industrial robot arms would learn to imitate motions that humans make while wearing a VR headset and controllers. If you want a robot to snake wiring through a car door, do the motion in VR, and the robot will attempt it in real life.
“What’s different about imitation learning is that you can actually teach a robot specifically what you want it to do,” Abbeel tells Quartz.
The Embodied Intelligence team (Embodied Intelligence)
Previous research with robotic arms focused on learning to pick up new objects by repeating the action over and over again, learning what worked from the successful attempts and ignoring the unsuccessful ones—a subset of AI called reinforcement learning. But that research took days of robotic “arm farms” attempting the task 24/7, Abbeel says. Embodied Intelligence takes the human motion first, which is guaranteed to be the correct motion, and then applies similar reinforcement learning after to make it more efficient.
Pivotally, the startup will focus on the movement of the arm itself, rather than the grabby bit at the end of it. Since robotic hands, called end effectors, don’t yet match the complexity or general usage of a human hand, Abbeel says that movement is relatively simple (grabber open or closed) compared to arranging the position and angle of the arm itself.
Plus, if it sells a few dozen VR headsets, maybe Mark Zuckerberg can recoup some of that $2 billion Oculus acquisition.