A group of former Google Deepmind engineers have a potential solution to robotics’ 100,000-year data gap: you.
Jonathan Scholz, the CEO of Reimagine Robotics, told Business Insider that the startup plans to build robots that are actually useful by turning its customers into robot teachers.
Scholz, who previously led DeepMind’s applied robotics team, founded Reimagine last year with former Google colleagues Oleg Sushkov, Akhil Raju and Misha Denil.
The London and Sydney-based startup emerged from stealth earlier this month and is backed by venture capital firms Fly Ventures and Firstminute Capital.
In his first media interview since emerging from stealth, Scholz told Business Insider that Reimagine’s mission is to build robots that can be refined and taught in the field — a process known in the world of AI as “post-training” — stemmed from his experiences at Google.
Show don’t tell: Reimagine’s robots are designed to be grabbed and manipulated. Think robotics again
After three separate robotics projects he built at DeepMind failed to progress beyond the pilot phase, Scholz said he realized the problem wasn’t the underlying technology, but how it was deployed in the real world.
“It struck me that it wasn’t just a capacity issue,” he said. “What you really had to do to deploy robots there was to make them usable and adaptable by the workforce who actually understands the work.”
‘Monkey see, monkey do’
To bridge this gap, Scholz said Reimagine has adopted a “monkey see, monkey do” approach to learning for the company’s fleet of robotic arms and assemblers, some of which are mounted on surfaces while others move on wheeled platforms.
He said customers can teach robots new behaviors by showing them how to perform a task, watching them try, and then correcting them by physically manipulating the robot arm.
This approach aims to overcome a major challenge for robotics companies that has hindered real-world adoption: a shortage of training data.
Unlike large language models like OpenAI’s GPT, which are trained on a vast corpus of online text, there is a relative lack of real-world data to train the AI systems that underlie robots.
Scholz called the ‘100,000 year data gap’ a term coined by roboticist from UC Berkeley Ken Goudberg. The largest reported data set on robot training contains approximately one year of experience. By comparison, Goldberg estimates that it would take a human about 100,000 years to read and view all the text and images used to train leading AI models.
Despite this gap, there is interest in robotics exploded in recent years as investors bet the field is approaching its own ‘ChatGPT moment’.
Much of this hype is focused on humanoid robots, with Tesla and robotics startups like Figure and 1X all producing impressive demos of their bipedal bots and engaging AI models capable of generalizing across a wide range of tasks.
‘An expensive paperweight’
Reimagine is working with several manufacturing companies to test its robotic arms. In one implementation, with a company that extracts critical materials from used hard drives, the startup says it reduced the time it took to teach a robot a new task from one day to ten minutes.
Ultimately, Reimagine envisions a “downstream economy” of robot teachers and tutors. Think robotics again
Scholz said employees working with the robots quickly adopted the platform and started working with it, coming up with their own use cases and training Reimagine’s robots to perform new tasks. He added that this kind of hands-on supervision would be necessary to move robots beyond flashy demos real working population.
“If they have to call the robotics guys from whiz team, wherever they are, to fly it in and fix it, they’re just not going to be able to do that. It’s going to be an expensive paperweight,” Scholz said.
In the long term, Scholz also expects the industry will see the emergence of a “downstream economy” made up of robot trainers who will “bridge the gap” between manufacturers and factories by teaching robots to perform specific tasks and solve problems past any roadblocks.
“When I worked at DeepMind, customers had to call us and we would fly out and fix the robots. From their perspective, it would be great if they didn’t have to make the calls,” he said.
