Defined deep reinforcement learning for robotics
Sergey Levine
Profile
Sergey Levine is the reason robots can learn. As a professor of EECS at UC Berkeley and a co-founder of Physical Intelligence, he sits at the exact intersection where deep learning stopped being a vision-and-text trick and started controlling motors. His lab — RAIL (Robotic AI & Learning) — has shipped a steady stream of the algorithms now used by basically every team trying to make robots behave intelligently from data instead of from hand-coded rules.
He came up under Pieter Abbeel at Stanford, then planted a flag at Berkeley in 2016. His 2013 Guided Policy Search and 2015 End-to-End Training of Deep Visuomotor Policies (with Chelsea Finn and Trevor Darrell) showed something most roboticists thought was years away: a single neural network mapping raw pixels straight to motor torques, trained end-to-end. Later work with Google on QT-Opt put that idea into a warehouse full of arms learning to grasp from millions of trials. If you read the RT-1 / RT-2 papers and wondered who built the foundations underneath, a lot of it traces back to him.
In 2024 he co-founded Physical Intelligence with Karol Hausman and others — a startup chasing a foundation model for robots the way OpenAI chased one for language. Their π0 and π0.5 policies pair a vision-language backbone with a diffusion-based action expert, trained on tens of thousands of hours of cross-embodiment data. The company has raised over $1B at a ~$5.6B valuation and recently demoed robots doing tasks they were never explicitly trained on. He still teaches and runs the lab in parallel.
For developers learning AI, Levine is unusually accessible: his CS285 lectures are the deep RL course that everyone outside Berkeley also takes, free on YouTube. If you want to actually understand policy gradients, Q-learning, model-based RL, and offline RL from someone who built half of it, this is the source.
Key Articles & Papers
Guided Policy Search End-to-End Training of Deep Visuomotor Policies QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation Soft Actor-Critic Offline Reinforcement Learning: Tutorial, Review, and Perspectives How to Train Your Robot with Deep Reinforcement Learning: Lessons We Have Learned π0: A Vision-Language-Action Flow Model for General Robot Control π0.5: A VLA with Open-World GeneralizationVideos
Spotify Podcasts
![Sergey Levine - Building LLMs for the Physical World - [Invest Like the Best, EP.465]](/images/people/sergey-levine/spotify/5C11GE9OnpmZvYslyu9GRC_hu_64e80ea1b8afda4.webp)
