Sergey Levine

UC Berkeley Professor, Physical Intelligence Co-founder, Robot RL Expert

Sergey Levine

Home > People > Sergey Levine

Profile

FieldDetails
Current PositionAssociate Professor, UC Berkeley
LabRAIL Lab (Robotic AI & Learning Lab)
CompanyPhysical Intelligence Co-founder
PreviousGoogle Brain (2016-2021)
PhDStanford University

Key Contributions

  • RT Series Key Figure: Participated in RT-1, RT-2 development, established VLA paradigm
  • Robot Reinforcement Learning Pioneer: Developed RL methods that work on real robots
  • Open X-Embodiment: Led collaboration across 33 research labs
  • OpenVLA, Octo: Developed open-source VLA models
  • Physical Intelligence Founding: Co-founded general-purpose robot AI company

Research Timeline

PhD & Postdoc (2009-2016)

Stanford -> UC Berkeley Postdoc

YearWorkImpact
2013Guided Policy SearchFoundation for model-based RL
2015End-to-End Visuomotor PoliciesDirect image-to-action learning
2016Deep RL for Robotic ManipulationPractical robot RL

Google Brain (2016-2021)

Google Robotics Research

YearWorkImpact
2017QT-OptLarge-scale robot grasping
2018Soft Actor-Critic (SAC)Most widely used off-policy RL
2020Offline RL SurveyEstablished offline RL
2021Decision TransformerRL as sequence modeling

UC Berkeley + Google (2021-2024)

RT Series, Open-source VLA

YearWorkImpact
2022RT-1First large-scale Robotics Transformer
2023RT-2First VLA, Action as Language
2023RT-XOpen X-Embodiment, 33 lab collaboration
2024Octo93M open-source generalist policy
2024OpenVLA7B open-source VLA

Physical Intelligence (2024-present)

Co-founded Physical Intelligence, pi0 development

Major Publications

Reinforcement Learning

  • SAC (Soft Actor-Critic) - Most widely used off-policy RL
  • CQL (Conservative Q-Learning) - Core offline RL method
  • Decision Transformer - RL as sequence modeling
  • Offline RL Tutorial - Established the field

Robot Learning

  • RT-1 (2022) - Robotics Transformer
  • RT-2 (2023) - Vision-Language-Action
  • RT-X (2023) - Open X-Embodiment
  • Octo (2024) - Open-source generalist policy
  • OpenVLA (2024) - 7B open-source VLA

End-to-End Learning

  • End-to-End Training of Deep Visuomotor Policies (2016)
  • QT-Opt (2018) - Large-scale robot grasping

Key Ideas

SAC (Soft Actor-Critic, 2018)

Core: Maximum entropy RL - maximize reward + maximize policy entropy

J(pi) = Sum E[r(st,at) + alpha*H(pi(.|st))]

Impact:

  • Currently the most widely used continuous control RL algorithm
  • Standard across robotics, games, and simulation

RT-2 & Action as Language (2023)

Core: Represent robot actions as text tokens to integrate with VLM

[Image + Language instruction] -> VLM -> [Action tokens] -> Robot control

Impact:

  • Beginning of VLA paradigm
  • All subsequent VLA models adopted this approach

Open X-Embodiment (2023)

  • Collaboration across 33 research labs
  • 22 robot types, 1M+ episodes
  • Democratized research with open-source dataset

Philosophy & Direction

Research Philosophy

“The key to robot learning is data. More data, more diverse data is the key to generalization.”

Research Direction Evolution

  1. 2009-2015: Model-based RL, trajectory optimization
  2. 2015-2018: End-to-end deep RL
  3. 2018-2021: Off-policy RL, offline RL
  4. 2021-2023: Large-scale robot learning, foundation models
  5. 2023-present: VLA, general-purpose robot AI

Students & Mentees

Sergey Levine lab alumni:

  • Chelsea Finn (Stanford Professor)
  • Aviral Kumar (Google DeepMind)
  • Ilya Kostrikov (Google DeepMind)
  • Numerous robotics/RL researchers

Awards & Recognition

  • IEEE RAS Early Career Award
  • Sloan Research Fellowship
  • NSF CAREER Award
  • ICML Best Paper (multiple)

See Also