Sergey Levine

Career

Levine received a PhD in computer science from Stanford University in 2014 and joined the faculty of UC Berkeley's Department of Electrical Engineering and Computer Sciences in fall 2016. Levine's research group is the Robotic Artificial Intelligence and Learning Lab (RAIL).¹

At Berkeley, Levine teaches CS 185/285 Deep Reinforcement Learning, whose lecture recordings are shared publicly. Levine's homepage also collects talks on topics such as real-world reinforcement learning, offline reinforcement learning and planning with reinforcement learning.⁵¹

Research

Levine's work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning. Early contributions include Trust Region Policy Optimization (2015) and end-to-end training of deep visuomotor policies that map raw image observations directly to torques at a robot's motors (2015), and Levine co-authored Soft Actor-Critic (2018), an off-policy maximum-entropy algorithm.¹⁴⁶⁷

Levine has also worked on offline reinforcement learning, writing a 2020 tutorial and review of the area with Aviral Kumar, George Tucker and Justin Fu and co-authoring Conservative Q-Learning. More recently, Levine co-authored π0, a vision-language-action flow model for general robot control whose weights are published in Physical Intelligence's openpi repository.⁸⁹²³

Work and contributions

• π0 (openpi) Open-source code

  Co-author of "π0: A Vision-Language-Action Flow Model for General Robot Control" (2024). The Apache-2.0 openpi repository, published by the Physical Intelligence team, provides the π0 model with base checkpoints for fine-tuning.²³

• CS 185/285 Deep Reinforcement Learning Published research

  Listed as the instructor of the UC Berkeley deep reinforcement learning course, whose site links lecture recordings and materials from previous offerings.⁵

• Trust Region Policy Optimization (TRPO) Published research

  Co-author with John Schulman, Philipp Moritz, Michael I. Jordan and Pieter Abbeel of the 2015 paper describing the TRPO policy optimization algorithm.⁴

Timeline

1. 2014Completed a PhD in computer science at Stanford University.¹
2. Feb 2015Co-authored "Trust Region Policy Optimization".⁴
3. Apr 2015Posted "End-to-End Training of Deep Visuomotor Policies" as first author.⁶
4. 2016Joined the faculty of UC Berkeley's Department of Electrical Engineering and Computer Sciences.¹
5. Jan 2018Co-authored Soft Actor-Critic, an off-policy maximum-entropy deep reinforcement learning algorithm.⁷
6. May 2020Posted a tutorial and review of offline reinforcement learning as first author.⁸
7. Jun 2020Co-authored "Conservative Q-Learning for Offline Reinforcement Learning".⁹
8. Oct 2024Co-authored the π0 vision-language-action flow model paper.²

Where they work

• UC Berkeley 27 people in the wiki

Related people

• Pieter AbbeelAlso at UC Berkeley
• Anca DraganAlso at UC Berkeley
• Aravind SrinivasAlso at UC Berkeley
• Jitendra MalikAlso at UC Berkeley
• Ken GoldbergAlso at UC Berkeley
• Deepak PathakRobotics · Reinforcement learning
• Bob McGrewRobotics · Reinforcement learning
• Jonathan HurstRobotics · Reinforcement learning
• Tuomas HaarnojaReinforcement learning · Robotics
• Justin FuReinforcement learning · Robotics