Can tech companies learn to love cheaper AI models?
Jay Li doesn’t recommend getting sued by Tesla if you’re trying to get a startup off the ground. But he does think his company, Proception, might be better off for having endured the experience. “I think it’s kind of like a resilience test, or pressure test,” he told TechCrunch in an exclusive interview. “People say that what doesn’t kill you makes you stronger, right?” Li, who was a technical lead on Tesla’s Optimus humanoid robot program, was accused by his former employer last year of absconding with trade secrets to start Proception. But after months of trading legal blows, he finally reached a settlement with Tesla, which dismissed the lawsuit earlier this month. (Tesla did not respond to a request for comment.) Now Li is free to tackle what he thinks is an even harder problem: making robot hands work like a human’s. To help do that, Proception announced Monday that it has raised an $11 million seed round led by First Round Capital, with contributions from Y Combinator and early stage fund BoxGroup. Proception also announced Monday that it is shipping the first batch of its “high-dexterity robotic hand” to “researchers and robotics companies,” while opening up to wider orders. The goal, Li said, is to become the top hand supplier to other companies that don’t want to spend the time or resources developing what’s known in the industry as “dextrous manipulation.” While there’s been an avalanche of money and attention rushing into the world of robotics, Li believes not enough of that has gone to making robotic hands truly mimic a human’s hands. One of the loudest voices talking about this challenge has actually been his old boss, Tesla CEO Elon Musk, who has said robot hands are one of the biggest engineering problems yet to be solved. While Musk has maintained that Optimus robots could start working in factories in a matter of years, the consensus view is that making robotic hands equivalent to a human’s is still many years away. Kevin Lynch, the director of Northwestern University’s Center for Robotics and Biosystems, told the Wall Sreet Journal last year that his team believes it will be a decade until they are “functional and useful and able to do some of the things that humans do.” Li thinks Proception can do it much faster, in large part because of how they’re collecting data. Most companies training humanoid robots right now are using teleoperators to train their systems. A human wearing a virtual reality headset is able to see what a robot sees and manipulate what’s in front of that robot, then the robot can learn from the commands given by the human. A big drawback to this approach, according to Li, is that the teleoperator is not receiving feedback from the objects the robot is touching. This approach is also limited to the number of robots a company has available at any given moment, Li said. Proception’s solution is a glove laden with sensors. With human testers wearing the gloves (and a headset), Proception and its customers can capture “human hand interaction data without requiring a robot in the loop,” according to Proception’s press release. This same glove also goes on the hand Proception is developing, acting as its sensor-packed “skin.” The hand has 22 degrees of freedom and multiple joints per finger to enable a “wide range of dexterous motions,” according to Proception. Li said this approach will also let Proception and its customers gather finer, more task-specific data that can allow its robotic hands to more accurately resemble a human’s. He also thinks it is better suited to scale up. “You need both hardware and data, and those need to come hand-in-hand to get [dextrous manipulation] to work. A lot of companies solely focus on hardware, or like hardware plus non-scalable data [collection],” he said. “We’re working on this highly dexterous hardware plus highly scalable data. We believe that’s a key combination to solve this problem.” First Round partner Bill Trenchard, who led the inve
