Sergey Levine

In 2009 Levine received a BS and an MS in computer science from Stanford UniversityStanford UniversityStanford University is a research university in California founded by Leland and Jane Stanford in 1885 and opened in 1891.Open the full entry.⁠[1]

In 2014 he completed his PhD at Stanford University with Vladlen Koltun, developing in his thesis a guided policy search algorithm for learning locomotion policies; he then became a postdoctoral researcher with Pieter Abbeel at UC Berkeley.⁠[2]

In August 2015 the University of Washington College of Engineering announced that he would join Computer Science & Engineering in the spring as an assistant professor, noting his graduate fellowships from NVIDIANVIDIANVIDIA is an American semiconductor and computing technology company that provides graphics processing units, artificial intelligence infrastructure, networking interconnects, and computing software, and is listed on Nasdaq under the ticker NVDA.Open the full entry and the National Science Foundation.⁠[3]

In 2016 he, Chelsea Finn, Trevor Darrell and Pieter Abbeel published “End-to-End Training of Deep Visuomotor Policies” in volume 17 of the Journal of Machine Learning Research, asking whether training perception and control jointly can outperform training them separately.⁠[4]

On March 8, 2016, as a GoogleGoogleGoogle is an American technology company founded by Larry Page and Sergey Brin and incorporated in 1998. It became an Alphabet subsidiary in 2015. Its businesses include internet search, advertising, software platforms, cloud computing, and artificial intelligence.Open the full entry research scientist, he described robotic grasping research in which, after more than 800,000 grasp attempts—about 3,000 robot-hours of practice—the system showed reactive behaviors such as correcting its motions in real time, akin to hand-eye coordination.⁠[5]

In fall 2016 he joined the faculty of UC Berkeley’s Department of Electrical Engineering and Computer Sciences, focusing on machine learning for decision making and control, especially deep learning and reinforcement learning algorithms, and leading the Robotic Artificial Intelligence and Learning Lab (RAIL).⁠[6]

In 2018 he, Tuomas Haarnoja, Aurick Zhou and Pieter Abbeel presented Soft Actor-Critic at the 35th International Conference on Machine Learning, an off-policy actor-critic deep RL algorithm based on maximum entropy reinforcement learning that achieved state-of-the-art, stable results on continuous control benchmarks.⁠[7]

In February 2019, then an assistant professor, he was named a 2019 Sloan Research Fellow; winners receive $70,000 over two years toward research.⁠[8]

In 2020 he, Aviral Kumar, Aurick Zhou and George Tucker published Conservative Q-Learning (CQL) at NeurIPS 2020, which learns a conservative Q-function to address distributional shift in offline reinforcement learning and often reached 2–5 times higher final returns than existing offline methods.⁠[9]

On July 14, 2020, he appeared on episode 108 of the Lex FridmanLex FridmanLex Fridman is a Soviet-born, US-raised computer scientist and podcast host who works as an MIT research scientist on human-AI interaction and autonomous driving. His Lex Fridman Podcast is known for hours-long conversations with guests including heads of state and technology leaders.Open the full entry Podcast, discussing the gap between robots and humans, end-to-end learning in robotics, reinforcement learning, simulation, reward functions and Rich Sutton’s “Bitter lesson.”⁠[10]

In March 2024 Physical Intelligence, which he co-founded with Karol Hausman, Chelsea Finn, Brian Ichter and Lachy Groom, emerged from stealth with $70 million in seed funding led by Thrive Capital, aiming to build foundation models and learning algorithms for a wide range of robots. He said, “There are enough signs that the biggest obstacles to use robots in the real world are now solvable.”⁠[11]

On October 31, 2024, Physical Intelligence released π0, a general-purpose robot foundation model, with him among the authors. The company said the model spans images, text and actions, directly outputs low-level motor commands and can control a variety of robots.⁠[12]

On November 4, 2024, Physical Intelligence confirmed raising $400 million at a $2.4 billion post-money valuation, with investors including Jeff BezosJeff BezosJeff Bezos is an American entrepreneur and investor who founded Amazon in 1994, led it as chief executive for many years, and became its executive chair in 2021. He also founded the spaceflight company Blue Origin, bought The Washington Post in 2013, and set up philanthropic funds including the Bezos Earth Fund.Open the full entry, OpenAIOpenAIOpenAI is an artificial intelligence research and product organization founded in 2015 that develops the GPT model series, ChatGPT, Codex and a developer API, and conducts research on image, speech, video and intelligent agents. After its 2025 restructuring, the nonprofit OpenAI Foundation continues to control the for-profit OpenAI Group PBC. Its growth has involved large-scale fundraising and compute partnerships, and has also been accompanied by controversies over governance, copyright, privacy and AI safety.Open the full entry, Thrive Capital and Lux CapitalLux CapitalLux Capital is a U.S. venture capital firm founded in 2000 by Peter Hébert, Josh Wolfe and Robert Paull, with offices in New York and Silicon Valley. Focused on emerging science and technology, the firm backs companies in fields including life sciences, artificial intelligence, robotics, manufacturing, aerospace and defense through fund investments and by helping to found companies.Open the full entry.⁠[13]

In January 2025, as an associate professor of electrical engineering and computer sciences, he received the Presidential Early Career Award for Scientists and Engineers (PECASE), the U.S. government’s highest honor for early-career scientists and engineers.⁠[14]

On February 4, 2025, Physical Intelligence released π0’s code and model weights in its openpi repository so others could fine-tune it for their own robots and tasks.⁠[15]

On April 22, 2025, Physical Intelligence released π0.5, a vision-language-action model co-authored by him that targets generalization to new settings and new objects.⁠[16]

On November 17, 2025, Physical Intelligence released π*0.6, co-authored by him, describing how its generalist policies are trained with reinforcement learning to improve success rates and throughput from experience.⁠[17]

On November 20, people with knowledge of the matter said Physical Intelligence had raised $600 million in a round led by Alphabet’s CapitalG, valuing the company at $5.6 billion.⁠[18]

On October 13, 2026, Physical Intelligence launched the Pi API beta, letting users upload data, fine-tune models on compute it manages and run inference through a hosted gateway.⁠[19]

As of October 2026 he is an associate professor in UC Berkeley’s Department of Electrical Engineering and Computer Sciences, holding the Harvey E. Wagner Presidential Chair in Artificial Intelligence.⁠[1]

Personal website: Public link: https://people.eecs.berkeley.edu/~svlevine/ (opens in a new window)⁠[1]

X: Public account: https://x.com/svlevine (opens in a new window)⁠[6]

UC Berkeley: Official website: https://www.berkeley.edu/ (opens in a new window)⁠[1]