Profile
| Field | Details |
|---|---|
| Current Position | Associate Professor, UC Berkeley |
| Lab | RAIL Lab (Robotic AI & Learning Lab) |
| Company | Physical Intelligence Co-founder |
| Previous | Google Research (2016-2024, concurrent) |
| PhD | Stanford 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
| Year | Work | Impact |
|---|---|---|
| 2013 | Guided Policy Search | Foundation for model-based RL |
| 2015 | End-to-End Visuomotor Policies | Direct image-to-action learning |
| 2016 | Deep RL for Robotic Manipulation | Practical robot RL |
UC Berkeley + Google (2016-2024)
UC Berkeley Faculty & Google Research Collaboration
| Year | Work | Impact |
|---|---|---|
| 2018 | QT-Opt | Large-scale robot grasping |
| 2018 | Soft Actor-Critic (SAC) | Most widely used off-policy RL |
| 2020 | Offline RL Survey | Established offline RL |
| 2021 | Trajectory Transformer | RL as sequence modeling |
RT Series & VLA (2022-2024)
RT Series, Open-source VLA
| Year | Work | Impact |
|---|---|---|
| 2022 | RT-1 | First large-scale Robotics Transformer |
| 2023 | RT-2 | First VLA, Action as Language |
| 2023 | RT-X | Open X-Embodiment, 33 lab collaboration |
| 2024 | Octo | 93M open-source generalist policy |
| 2024 | OpenVLA | 7B 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
- Trajectory Transformer - RL as sequence modeling (concurrent with Decision Transformer)
- 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 controlImpact:
- 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
- 2009-2015: Model-based RL, trajectory optimization
- 2015-2018: End-to-end deep RL
- 2018-2021: Off-policy RL, offline RL
- 2021-2023: Large-scale robot learning, foundation models
- 2023-present: VLA, general-purpose robot AI
Students & Mentees
Sergey Levine lab alumni:
- Aviral Kumar (CMU Professor, Google DeepMind Researcher)
- Justin Fu (Google DeepMind)
- Michael Janner (UC Berkeley)
- Numerous robotics/RL researchers
Key collaborators:
- Chelsea Finn (Stanford Professor, Physical Intelligence Co-founder)
- Pieter Abbeel (UC Berkeley Professor)
Awards & Recognition
- Presidential Early Career Award for Scientists and Engineers (PECASE)
- Sloan Research Fellowship (2019)
- CoRL Best Systems Paper (QT-Opt, 2018)
- Multiple conference Best Paper Awards