Train Your Robot as Easily as You Train Your Dog
Lux Capital invests in Embodied Intelligence
Factory robots are a marvel at doing dangerous, precise, repetitive tasks, such as welding together cars, shuttling metal sheets between giant stamping presses, and placing tiny chips on circuit boards. Yet somehow they are pathetically discombobulated when it comes to the most mundane tasks: everyday chores, such as folding towels or picking up around the house.
That is all about to change. Today Embodied Intelligence is announcing a $7 million seed round, including participation from Lux Capital. Embodied will use the cash infusion to further develop a technology that makes it possible to train a robot as easily as you train your dog. Pretty soon you will not need a room full of PhDs to teach a robot how to sort through a box full of wires or a conveyor belt lined with irregularly-shaped cookies.
Embodied’s cofounder and Chief Scientist, Pieter Abbeel, first showed up on my radar over half a decade ago when I was touring Willow Garage on Willow Road in Menlo Park. This nondescript building tucked away in a suburban neighborhood housed the best minds in computer vision and robotics. They had built the PR2 robot: a $100,000, two-armed, mobile robot armed with cameras and sensors. This army of PhDs had painstakingly programmed the PR2 to shoot pool and grab a beer. It was a neat trick, but the robots were dreadfully slow. I asked the robot geniuses which customer was doing the most cutting-edge work on training the robots to do interesting tasks at a speed close to human capability. Several answered: Professor Pieter Abbeel at UC Berkeley.
I quickly reached out to Pieter and asked to visit him in his lab. He accepted, and as he showed me around, he pointed out a PR2 robot standing next to a pile of towels. The robot sprang to life, and meticulously folded the laundry into a neat pile. It was something to see!
Pieter centered his research around reinforcement learning. This was underpinned by his seminal work with Andrew Ng at Stanford where he programmed a helicopter to do acrobatics by watching another helicopter do the same maneuvers. I wanted to fund a company around Pieter then, but the research was still too early, and the robots were still too expensive and too slow.
Fast-forward to 2017. I heard through the robot grapevine that Pieter was putting together a new company around training robots. Rumor was, he was able to get the robots to do simple tasks at a much quicker pace than before. Rather than being ten times slower than humans, robots could now keep up with their human counterparts! Most importantly, training the robots didn’trequire PhDs or state-of-the-art tools. Now, anyone skilled enough to perform the task could teach a robot using off-the-shelf electronics from Best Buy.
Embodied aims to create more jobs and prosperity for humanity by vastly improving the quality of human output with the aid of robots. I’m looking forward to the insights they gather from their efforts to rethink machines. If they are successful, I’m confident they will play an important role in accelerating humanity toward a brighter future.
