Sergey Levine is focused on a simple but powerful idea. Machines should not just follow instructions. They should learn how to get better at things on their own.
At the University of California, where he is an associate professor of computer science, Levine is building systems that do exactly that. Most traditional systems break the moment something unexpected happens but Levine’s work starts from the opposite assumption. The real world is messy, so machines must learn by trying, failing, and adjusting.
Since joining Berkeley in 2016, Sergey Levine has pushed to make machine learning work outside the lab, in real environments where unpredictability is the norm. In his lab, robots are assembling hardware, handling flexible objects and recovering from mistakes mid-task. In some cases, they can learn these skills in just one or two hours of real-world practice.
Importantly, Levine is less interested in what looks impressive and more focused on what works consistently. For him, a real milestone is a machine doing a task well enough that someone is willing to pay for it.
Alongside his academic work, Levine co-founded Physical Intelligence, a company focused on building software that can work across many types of robots. His work increasingly shapes how the field thinks about “embodied intelligence”— not as a distant ambition, but as a set of systems already beginning to function in the real world.
As a speaker, Sergey Levine stands out for his precision. He does not rely on sweeping claims or distant futures. Instead, he explains where the field actually is, what constraints still matter, and which signals indicate real progress. He connects technical ideas to economic reality, offering audiences a way to think clearly about automation without oversimplifying it.
