Professor research guide
Associate Professor · Electrical Engineering and Computer Sciences
Sergey Levine studies learning methods for autonomous agents at UC Berkeley and is a co-founder of Physical Intelligence, a separate company. This independent guide distinguishes academic RAIL results from Physical Intelligence research.
Official profile ↗Independent educational guide. Not affiliated with or endorsed by the universities, professors or laboratories described here. This collection reflects the material currently mapped on Socratic Learn.
Berkeley RAIL research studies robot learning from demonstrations, reward, prior datasets, and interaction: policies, reinforcement learning, manipulation, and generalization. The independent Levine primer also covers his co-founder work at Physical Intelligence, a separate company.
7 lessons · ~37 minutes
Explore the labYour recommended path
Research primer
7 lessons · ~37 minutes
A video-first introduction to robot policies, imitation and reinforcement learning, offline learning, goal-conditioned behavior, generative control, and robot foundation models. Academic RAIL research and Physical Intelligence company research are identified separately.
Explain behavioral cloning and compounding error under distribution shift.
Explain conservative value estimates and cautious offline-to-online improvement.
Explain cross-embodiment data and the subtask-action-observation hierarchy.
Describe the distinct contributions of demonstrations, interventions, and human reward labels.
Distinguish diffusion and continuous flow matching from language token prediction.
Explain why averaging distinct valid actions can be unsafe.
Explain how a goal changes the action chosen by one policy.
Explain how autonomous attempts expose policy-specific failures.
Compare learned visuomotor control with an engineered pipeline.
Explain why fixed-data learning cannot test an unfamiliar action.
Explain the roles of subgoals and temporally coherent action chunks.
Distinguish a demonstration from reward and explain expected return and Q-values.
Explain a policy and the closed perception-action loop.
Distinguish a VLM from a VLA and a company goal from demonstrated capability.
2016
Levine S, Finn C, Darrell T, Abbeel P
Why this matters: Berkeley / RAIL research. Joint perception and motor learning for evaluated robot tasks; not universal competence.
2018
Haarnoja T, Zhou A, Abbeel P, Levine S
Why this matters: Berkeley / RAIL research. Connects expected return, stochastic policies, and off-policy learning.
2020
Levine S, Kumar A, Tucker G, Fu J
Why this matters: Berkeley / RAIL research. Explains fixed datasets and the fundamental distribution-shift problem.
2020
Kumar A, Zhou A, Tucker G, Levine S
Why this matters: Berkeley / RAIL research. Introduces conservative values for poorly supported actions.
2020
Nair A, Gupta A, Dalal M, Levine S
Why this matters: Berkeley / RAIL research. Links learning from prior datasets to subsequent online improvement; venue remains unconfirmed in the supplied source.
2023
Park S, Ghosh D, Eysenbach B, Levine S
Why this matters: Berkeley / RAIL research. Uses latent subgoals to connect a high-level policy to goal-reaching behavior.
2023
Shah D, Sridhar A, Dashora N, Stachowicz K, Black K, Hirose N, Levine S
Why this matters: Berkeley / RAIL research. A visual-navigation foundation model evaluated across robot platforms.
2023
Zhao TZ, Kumar V, Levine S, Finn C
Why this matters: Berkeley / RAIL research. Multi-institution Berkeley, Stanford, and Meta work: chunking mitigates compounding errors without eliminating them.
2022
Janner M, Du Y, Tenenbaum JB, Levine S
Why this matters: Berkeley / RAIL research. Berkeley with MIT co-authors: plans trajectories by iterative denoising.
2024
Sridhar A, Shah D, Glossop C, Levine S
Why this matters: Berkeley / RAIL research. Uses a diffusion policy for goal-directed navigation and exploration.
2025
Park S, Li Q, Levine S
Why this matters: Berkeley / RAIL research. Combines a generative policy with value-guided learning.
2024
Black K, Brown N, Driess D, et al., Levine S, et al
Why this matters: Physical Intelligence research: pretrained vision-language knowledge plus a flow-matching action expert; evaluated tasks do not establish universal control.
2025
Physical Intelligence, Black K, Brown N, Darpinian J, et al
Why this matters: Physical Intelligence research: semantic subtasks and action chunks, with three unseen-home evaluations and explicit limitations.
2025
Shi LX, Ichter B, Equi M, Ke L, Pertsch K, Vuong Q, Tanner J, Walling A, Wang H, Fusai N, Li-Bell A, Driess D, Groom L, Levine S, Finn C
Why this matters: Physical Intelligence research with Stanford and Berkeley co-authors: hierarchical instruction following; not human-level reasoning.
2025
Physical Intelligence, Amin A, Aniceto R, Balakrishna A, Black K, et al
Why this matters: Physical Intelligence research: RECAP combines demonstrations, autonomous attempts, interventions, and human outcome labels; not fully autonomous.
2026
Physical Intelligence
Why this matters: Physical Intelligence current research: context-steered transfer examples with lower reported success on unseen tasks; transfer remains unresolved.