Sergey Levine

Sergey Levine is a computer scientist whose research focuses on machine learning for robots. He is a professor in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, a post he took up in 2016 ^[1], and a co-founder of Physical Intelligence ^[1][2].

Education

Levine took all three of his degrees — a bachelor's, a master's and a doctorate — in computer science at Stanford University ^[1]. His doctoral advisor was Vladlen Koltun ^[1].

Career

In 2015 Levine was a part-time research scientist at Google ^[1]. In the same year he joined the faculty of the Paul G. Allen School of Computer Science and Engineering at the University of Washington ^[1]. In 2016 he became a professor in the Department of Electrical Engineering and Computer Sciences at UC Berkeley ^[1].

Research

Levine's research covers deep reinforcement learning for robots, and three strands of that work are named in his research description: guided policy search, end-to-end visuomotor policy learning, and offline reinforcement learning from large-scale robot data ^[1].

Reinforcement learning is the general recipe in which a system tries actions, receives a score for how well it did, and adjusts its behaviour to earn better scores. Deep reinforcement learning does this with neural networks, which can represent complicated mappings from what a robot senses to what it should do. Applying the recipe to a physical robot is awkward, because every attempt takes real time and a badly chosen action can damage the hardware or the surroundings.

Guided policy search addresses that difficulty through its name: instead of searching directly for a neural-network controller by unguided trial and error, the search is steered by a simpler, more tractable source of guidance, so that the network ends up with behaviour that is easier to obtain by other means but still generalises when the robot encounters new situations.

End-to-end visuomotor policy learning removes the usual intermediate steps. In a conventional robotics stack, camera images are first turned into something else — object positions, for instance — by hand-built perception code, and a separate module then decides on motor commands. An end-to-end visuomotor policy instead maps images straight to actions inside one learned model, so that perception and control are trained together rather than designed separately and glued at an interface.

Offline reinforcement learning from large-scale robot data attacks a different bottleneck: collecting robot experience is slow and expensive, and a learning method that must keep trying new actions during training spends most of that effort on the robot itself. Offline methods learn from a fixed body of data gathered earlier, so that the same logged experience can be reused and learning can happen without fresh interaction.

Company

Levine is a co-founder of Physical Intelligence, and he is listed on the company's team page ^[1][2].

Place in the field

The three research strands share a common direction. Each replaces a piece of hand-engineered machinery — a hand-tuned controller, a hand-designed perception pipeline, or a training loop that depends on live trial and error — with learning from data. That is why the work is usually grouped under robot learning rather than under any single application, and why it connects the questions of control, perception and data collection that robotics normally treats separately.