Robots Get AI from Startup

　　In the next few months, industrial robots will learn how to do their jobs by watching humans, using software from a startup that debuts today. The neural-networking program from Embodied Intelligence also will let robots improve their performance over time.

　　The work marks a step toward a future in which robots will understand the visual world. Today, human experts typically train factory-floor robots to repeat motions in a relatively slow two-step process that sometimes requires humans writing custom software.

　　“Instead of programming each procedure, we demo it — it doesn’t require an expert … the robot learns from trial and error,” said Peter Chen, a co-founder and chief executive of the company.

　　“Our robot software is not restricted to fixed motions. Today, robots do the same mechanical tasks over and over. Our software gives robots the ability to really see through their cameras and make adjustments.”

　　In addition to training robots faster and more cheaply, the software also opens the door to teaching new tasks. For example, the system could teach a robot how to thread a wire through a mechanical part, something most computer-vision systems cannot do given the complexity of tracking and programming for a flexible object.

　　The startup uses virtual reality headsets to train robots. It currently uses the HTC Vive headset and its motion controller, although any VR headset will do.

　　“You see what the robot sees, you make decisions based on what the robot sees … and the robot imitates it,” he said.

　　Chen was one of three Berkeley researchers who published results earlier this year of experiments teaching robots 10 basic tasks using machine learning and a VR connection. “With a three-minute demo in VR, robots solved all tasks that previously might have required a PhD in writing algorithms,” he said.

　　The approach uses the same deep neural network techniques that web giants such as Google and Facebook use to recognize images and other tasks. VR demos act as the training, setting up neural network pathways or policies that the robots later refine by running inference tasks.

　　The company currently builds its own Linux x86 servers using up to eight high-end Nvidia GPUs for training and one for inference work.

　　“In the beginning, we will provide this as a service for users who come to us with their specs … that will help us perfect our platform,” he said. “At some point, we will license the software to systems integrators.”

　　Chen claims that most of the money that a factory spends on a robot goes to systems integrators who train them — as much as $90,000 of an average total of $150,000. “We are going after that $90,000,” he said.

　　Others agree that the brunt of the cost of a robot lies outside the base hardware, much of it in training.

　　Factories are expected to buy more than 300,000 robots this year, said Dan Kara, research director for robotics at market watcher ABI Research. He pegs the average cost of an industrial robot at $42,000 and an installed and trained system at $126,000, much of it in software development.

　　“Programming industrial robots is a difficult, costly, and time-consuming task,” said Kara in an email exchange. “Tools and techniques that simplify and speed robot-control programming are in high demand.”

　　Kara listed Fizyr, Osaro, and Preferred Networks as three other companies working on teaching industrial robots. Google and Brown University are among others doing research in the area.

　　Henrik I. Christensen, director of the Institute for Contextual Robotics at U.C. San Diego, said that PlusOne, Universal Robotics, and researchers in Seattle are also pursuing the area.

　　“There are quite a few groups trying to use machine learning for robotics,” said Christensen.

　　“The reality is that use of machine learning is still quite limited in this area,” said Chen. “The most common thing is using machine learning in inspection; many people are doing that.”

　　Chen and two Berkeley colleagues co-authored more than 180 papers in the field before they founded Embodied Intelligence. The trio worked together at OpenAI for about 18 months when they decided that there was a commercial gap they thought they could fill.

　　Other founders include Berkeley alums Pieter Abbeel and Rocky Duan, the startup’s chief scientist and chief technology officer, respectively. They were joined by Tianhao Zhang from Microsoft Research as a fourth co-founder.

　　The startup raised a $7 million seed round, which Chen said could take it through its first two years. Investors were led by Amplify Partners and include Lux Capital, 11.2 Capital, A.Capital, SV Angels, Rostrum Capital, and angel investors such as Lip-Bu Tan, chief executive of Cadence.