Chelsea Finn
Last reviewed
Sources
17 citations
Review status
Source-backed
Revision
v3 · 3,730 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
17 citations
Review status
Source-backed
Revision
v3 · 3,730 words
Add missing citations, update stale details, or suggest a clearer explanation.
Chelsea Finn (born October 8, 1992) is an American computer scientist, an assistant professor of computer science and electrical engineering at Stanford University, and a co-founder of the robotics company Physical Intelligence. She is best known for inventing Model-Agnostic Meta-Learning (MAML), a 2017 algorithm that trains a neural network so it can adapt to a brand-new task from only a few examples and a few gradient steps; the MAML paper has been cited roughly 19,955 times and is the most-cited work in gradient-based meta-learning. [5][16] Finn directs Stanford's IRIS Lab and is a leading figure in meta-learning, robot learning, and foundation models for robotic manipulation, with hardware and software platforms including ALOHA and Mobile ALOHA for low-cost bimanual manipulation. [1][10][11]
Her doctoral work at UC Berkeley, advised by Pieter Abbeel and Sergey Levine, introduced MAML, one of the most influential algorithms for few-shot learning in deep neural networks. [5][6] Her research, in her own words, lies "at the intersection of machine learning and robotic control, including topics such as end-to-end learning of visual perception and robotic manipulation skills, deep reinforcement learning of general skills from autonomously collected data, and meta-learning algorithms that can enable fast learning of new concepts and behaviors." [1] As of 2026 her publications had accumulated more than 134,000 citations with an h-index of 127 on Google Scholar, making her one of the most-cited researchers in machine learning. [16]
Finn directs the IRIS (Intelligence through Robotic Interaction at Scale) Lab at Stanford, which is affiliated with the Stanford AI Lab (SAIL). Her research includes end-to-end learning of visual perception and manipulation, deep reinforcement learning from autonomously collected data, imitation learning, and meta-learning algorithms that allow agents to acquire new skills from very small amounts of data. In March 2024 she co-founded Physical Intelligence in San Francisco with Karol Hausman, Sergey Levine, and several collaborators, where she contributes to the team that built the π₀ generalist robot policy. [12]
Finn has been recognized with the ACM Doctoral Dissertation Award, the Sloan Research Fellowship, the IEEE Robotics and Automation Society Early Academic Career Award, the Office of Naval Research Young Investigator Award, the Microsoft Research Faculty Fellowship, and the Presidential Early Career Award for Scientists and Engineers (PECASE). [7][8][9][15] Her papers, talks, and demonstrations have been covered widely in outlets such as The New York Times, Wired, Bloomberg, and TechCrunch, and the Mobile ALOHA video she released with her students in January 2024 became one of the most viral robotics demonstrations of that year. [14]
| Field | Details |
|---|---|
| Born | October 8, 1992 |
| Nationality | American |
| Education | B.S., Electrical Engineering and Computer Science, MIT (2014); Ph.D., Computer Science, UC Berkeley (2018) |
| Doctoral advisors | Pieter Abbeel and Sergey Levine |
| Doctoral thesis | Learning to Learn with Gradients (2018) |
| Known for | Model-Agnostic Meta-Learning (MAML), ALOHA, Mobile ALOHA, π₀, robot learning, foundation models for robotics |
| Institutions | Stanford University (since September 2019); Google Brain (research scientist, concurrent); Physical Intelligence (co-founder, 2024 to present) |
| Lab | IRIS (Intelligence through Robotic Interaction at Scale), Stanford |
| Notable awards | ACM Doctoral Dissertation Award (2019); Microsoft Research Faculty Fellowship (2020); Samsung AI Researcher of the Year (2020); ONR Young Investigator (2021); IEEE RAS Early Academic Career Award (2022); Sloan Research Fellowship (2023); PECASE (2025) |
Chelsea Brittany Finn was born on October 8, 1992. Her early interests in mathematics, engineering, and artificial intelligence led her to the Massachusetts Institute of Technology (MIT), where she enrolled as an undergraduate in the Department of Electrical Engineering and Computer Science (EECS). She graduated with a Bachelor of Science in EECS in 2014. [1][4]
During her undergraduate years she participated in MIT's SuperUROP program as a Qualcomm Undergraduate Research and Innovation Scholar, conducting research on text detection in natural images for accessibility applications aimed at the visually impaired. The project bridged her interests in computer vision, machine learning, and applied engineering and helped motivate her later focus on perception-driven autonomy.
In the fall of 2014 Finn began doctoral studies in computer science at the University of California, Berkeley (UC Berkeley), where she joined the Berkeley Artificial Intelligence Research (BAIR) Lab. She was jointly advised by Pieter Abbeel and Sergey Levine, two of the leading figures in deep reinforcement learning and robot learning. [4] She held a National Science Foundation Graduate Research Fellowship (NSF GRFP) during her PhD.
While at Berkeley she also held research internships at Google's robotics and Google Brain teams, where she worked on visuomotor policy learning, deep predictive models for robotic interaction, and meta-learning. Her doctoral research culminated in the dissertation Learning to Learn with Gradients, defended in August 2018, in which she developed a class of meta-learning algorithms, most notably MAML, that train neural networks to be easy to fine-tune to new tasks with only a few examples and a few gradient steps. [6] The dissertation received the Association for Computing Machinery (ACM) Doctoral Dissertation Award in 2019, with the citation recognizing her foundational contributions to meta-learning and robot learning. [7] She had earlier been honored at Berkeley with the C.V. and Daulat Ramamoorthy Distinguished Research Award (2017) and was a Berkeley EECS Rising Star participant.
In September 2019 Finn joined Stanford University as an assistant professor with a joint appointment in the Department of Computer Science and the Department of Electrical Engineering. [1] At Stanford she founded and directs the IRIS Lab (Intelligence through Robotic Interaction at Scale), which is affiliated with the Stanford AI Lab (SAIL) and the Stanford Machine Learning Group, and is also part of the Stanford Robotics Center. She holds the William George and Ida Mary Hoover Faculty Fellowship at Stanford and is a faculty affiliate of the Stanford Institute for Human-Centered Artificial Intelligence (HAI). [1]
The IRIS Lab studies how learning algorithms can allow robots and other embodied agents to acquire broadly intelligent behavior through interaction. [3] Research themes in the lab include end-to-end visuomotor learning for manipulation, deep reinforcement learning from autonomously collected data, large-scale imitation learning, language-conditioned manipulation, robot foundation models, and meta-learning. The lab collaborates extensively with the Stanford ML Group, including with the ILIAD lab led by Dorsa Sadigh, on shared projects in robot learning and human-robot interaction.
In addition to her Stanford appointment, Finn served concurrently as a part-time research scientist at Google Brain (later folded into Google DeepMind), continuing collaborations on large-scale robot learning and on multi-task and multi-robot data efforts. She has also served on the program committees and as an area chair for major machine learning and robotics conferences including NeurIPS, ICML, ICLR, the Conference on Robot Learning (CoRL), and the IEEE International Conference on Robotics and Automation (ICRA).
Alongside her Stanford position, Finn has held a research scientist role at Google Brain. The arrangement, common among Stanford AI faculty, allowed her to continue working on large-scale robot learning datasets and projects, including contributions to the Open X-Embodiment effort that pooled robot manipulation data across more than twenty institutions. [13]
In March 2024 Finn co-founded Physical Intelligence, a San Francisco startup developing general-purpose foundation models for robots. [12] Her co-founders include Karol Hausman (CEO, previously a Staff Research Scientist at Google DeepMind and adjunct professor at Stanford), Sergey Levine (chief scientist), Brian Ichter, Lachy Groom, Quan Vuong, Suraj Nair, and Adnan Esmail. Physical Intelligence raised a $400 million Series A in November 2024 at a $2.4 billion post-money valuation, with investors including Jeff Bezos, Thrive Capital, Lux Capital, OpenAI, Sequoia Capital, Khosla Ventures, Redpoint Ventures, Bond Capital, and CapitalG. [17] Finn contributes to the company's research while continuing in her Stanford faculty role.
Finn's research is unified by a long-term goal: building robots that can learn broadly intelligent behavior through interaction with the physical world. [1] Her contributions span learning algorithms, hardware platforms, and large-scale data efforts.
The paper for which Finn is most widely known, "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks" (Finn, Abbeel, Levine), was published at the 34th International Conference on Machine Learning (ICML) in 2017. [5] MAML proposes a meta-learning algorithm that is compatible with any model trained by gradient descent and that can be applied to supervised classification, regression, and reinforcement learning. As the paper states, "the parameters of the model are explicitly trained such that a small number of gradient steps with a small amount of training data from a new task will produce good generalization performance on that task." [5] The key idea is to learn an initialization for a neural network's parameters such that a small number of gradient updates on a new task, with only a few examples, yields good generalization. The original paper has accumulated roughly 19,955 citations as of 2026 and remains the most widely cited algorithm in the gradient-based meta-learning literature, motivating dozens of follow-ups including first-order MAML, Reptile, ANIL, and probabilistic MAML. [16]
Finn extended MAML to robotic skills with works including "One-Shot Visual Imitation Learning via Meta-Learning" (Finn, Yu, Zhang, Abbeel, Levine, CoRL 2017), in which a robot learns to imitate a manipulation skill from a single human or robot video demonstration. With Tianhe Yu and others she developed Domain-Adaptive Meta-Learning (DAML), which adapts a meta-learned policy to a new visual context using a single demonstration in a new domain. These works helped establish meta-learning as a practical tool for sample-efficient robot learning rather than only a benchmark technique.
Finn co-authored "End-to-End Training of Deep Visuomotor Policies" (Levine, Finn, Darrell, Abbeel, JMLR 2016), one of the early demonstrations that convolutional neural network policies could be trained end-to-end to map raw camera pixels to torque commands for a real robot; the paper has been cited more than 5,000 times. [16] Related work on "Unsupervised Learning for Physical Interaction through Video Prediction" (Finn, Goodfellow, Levine, NeurIPS 2016) introduced video prediction models that learn from unlabeled robot interaction data, enabling planning by imagining future frames.
Finn's group at Stanford has been a leading contributor to offline reinforcement learning for robotics, including conservative Q-learning extensions, Implicit Q-Learning variants, and large-scale offline pretraining followed by online fine-tuning. Her group has also been a central contributor to the Open X-Embodiment dataset and to robot foundation models trained across many embodiments and tasks, including the RT-1 and RT-2 lines of work in collaboration with Google DeepMind. [13]
In April 2023 her student Tony Z. Zhao led the release of ALOHA, the "A Low-cost Open-source Hardware system for bimanual teleoperation," together with the imitation learning algorithm Action Chunking with Transformers (ACT). The accompanying paper, "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware" (Zhao, Kumar, Levine, Finn), was presented at Robotics: Science and Systems (RSS) 2023. [10] ALOHA enables fine-grained bimanual tasks such as threading a zip tie or assembling a chain at a hardware cost of under $20,000, and ACT predicts chunks of future actions at once instead of single actions, reaching 80 to 90 percent success on contact-rich tasks from just 10 minutes of demonstration data while dramatically reducing the compounding errors typical of behavior cloning. [10]
In January 2024 Finn's students Zipeng Fu and Tony Z. Zhao released Mobile ALOHA, an extension of ALOHA that adds a wheeled mobile base and whole-body teleoperation interface. The accompanying paper, "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation" (Fu, Zhao, Finn), was presented at the Conference on Robot Learning (CoRL) in 2024. [11] Using 50 demonstrations per task, co-training on static ALOHA data alongside mobile data increased success rates by up to 90 percent, letting the roughly $32,000 system autonomously sauté and serve shrimp, open a wall cabinet to store heavy pots, call and enter an elevator, and rinse a pan at a kitchen faucet. [11][14] The release video became one of the most widely shared robotics demonstrations of 2024.
In 2024 Finn was a co-author on two prominent vision-language-action models. π₀, released by Physical Intelligence on October 31, 2024, is a generalist robot policy that augments a pretrained vision-language model with continuous action outputs via flow matching, enabling high-frequency dexterous control across diverse robot platforms and tasks such as folding laundry, bussing tables, and assembling boxes; Physical Intelligence describes it as "the most capable and dexterous generalist robot policy to date" and open-sourced the model on February 4, 2025. [12] OpenVLA, released in June 2024 (Kim, Pertsch, Karamcheti, et al., with Finn as senior author), is a 7-billion-parameter open-source vision-language-action model trained on 970,000 robot demonstrations from the Open X-Embodiment dataset; it outperformed the closed 55-billion-parameter RT-2-X model by 16.5 percentage points in absolute task success across 29 tasks while using roughly 7 times fewer parameters. [13]
Finn's group has also been active in language-conditioned manipulation, in robot learning from internet videos, in detecting machine-generated text (the DetectGPT line of work led by her student Eric Mitchell), and in scaling laws and evaluation methodology for robotic foundation models. She is also a co-author on the widely cited 2021 report "On the Opportunities and Risks of Foundation Models" and on "Direct Preference Optimization" (DPO), each of which has been cited more than 10,000 times. [16]
Physical Intelligence, often abbreviated "π," was founded in San Francisco in March 2024 to build foundation models that can, in the company's words, "control any robot to do any task." [12] Its founding team brought together researchers from Stanford, UC Berkeley, and Google DeepMind. Karol Hausman serves as CEO. Sergey Levine is chief scientist. Other founders include Brian Ichter, Lachy Groom, Quan Vuong, Suraj Nair, Adnan Esmail, and Chelsea Finn. [17]
In November 2024 the company announced a $400 million Series A round at a $2.4 billion post-money valuation, led by Jeff Bezos and Thrive Capital, with participation from OpenAI, Lux Capital, Khosla Ventures, Redpoint Ventures, Sequoia Capital, Bond Capital, and CapitalG. [17] The company's first publicly released model, π₀, was unveiled on October 31, 2024, open-sourced on February 4, 2025, and was followed by π₀.5 and other extensions. [12] Subsequent funding rounds reportedly increased the company's valuation substantially in 2025.
Finn participates in the company as a co-founder and member of the engineering and research team while continuing her tenure-track role at Stanford.
The following table lists a selection of Finn's most cited and most influential publications. Citation counts are approximate Google Scholar numbers and continue to grow. [16]
| Year | Title | Venue | Role | Approx. citations |
|---|---|---|---|---|
| 2016 | End-to-End Training of Deep Visuomotor Policies | JMLR | Co-author with Levine, Darrell, Abbeel | 5,100 |
| 2016 | Unsupervised Learning for Physical Interaction through Video Prediction | NeurIPS | First author with Goodfellow, Levine | - |
| 2017 | Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks | ICML | First author with Abbeel, Levine | 19,955 |
| 2017 | One-Shot Visual Imitation Learning via Meta-Learning | CoRL | First author | - |
| 2018 | Learning to Learn with Gradients (PhD dissertation) | UC Berkeley | Sole author | - |
| 2020 | Conservative Q-Learning for Offline Reinforcement Learning | NeurIPS | Co-author | - |
| 2021 | On the Opportunities and Risks of Foundation Models | Report | Co-author | 10,600 |
| 2023 | Direct Preference Optimization (DPO) | NeurIPS | Co-author with Rafailov et al. | 10,400 |
| 2023 | Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ALOHA / ACT) | RSS | Senior author with Zhao, Kumar, Levine | - |
| 2023 | Open X-Embodiment: Robotic Learning Datasets and RT-X Models | ICRA 2024 | Co-author with multi-institution team | - |
| 2024 | Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation | CoRL | Senior author with Fu, Zhao | - |
| 2024 | OpenVLA: An Open-Source Vision-Language-Action Model | CoRL | Senior author with Kim, Pertsch, others | - |
| 2024 | π₀: A Vision-Language-Action Flow Model for General Robot Control | Physical Intelligence preprint | Co-author | - |
Finn teaches and co-teaches several courses at Stanford that are widely followed inside and outside the university. Her CS 330 course, in particular, has been recorded and posted publicly, and the YouTube playlists for the course are commonly used as canonical introductions to deep multi-task and meta-learning.
| Course | Title | Notes |
|---|---|---|
| CS 330 | Deep Multi-Task and Meta Learning | Created and taught annually since Fall 2019 |
| CS 224R | Deep Reinforcement Learning | Co-taught at Stanford from Spring 2023 |
| CS 221 | Artificial Intelligence: Principles and Techniques | Co-taught in Spring 2020 and Spring 2021 |
She has also given invited lectures and tutorials in venues such as the Deep Learning Indaba, the MIT EECS Rising Stars workshop, the NeurIPS meta-learning tutorials, and several summer schools on robot learning.
Finn has advised a growing group of doctoral students and postdoctoral researchers at Stanford. The following table lists several of her best-known advisees and their primary lines of work.
| Advisee | Primary work | Subsequent role |
|---|---|---|
| Tony Z. Zhao | ALOHA, Mobile ALOHA, ACT | Robotics researcher; co-founded Generalist AI |
| Zipeng Fu | Mobile ALOHA, legged locomotion, humanoid teleoperation | Stanford PhD researcher |
| Eric Mitchell | DetectGPT, model editing, RLHF alternatives | Co-founder of Goodfire AI |
| Annie Xie | Reinforcement learning for adaptation in dynamic environments | Research scientist, Google DeepMind |
| Suneel Belkhale | Language-conditioned manipulation, hierarchical policies | Stanford PhD researcher |
| Karl Pertsch | Robot foundation models, OpenVLA | Postdoctoral researcher, Stanford and UC Berkeley |
| Suraj Nair | Visual representations for manipulation, R3M | Co-founder, Physical Intelligence |
Finn has received a series of recognitions for her research and dissertation work. [7][8][9][15]
| Year | Award |
|---|---|
| 2014 | NSF Graduate Research Fellowship |
| 2017 | UC Berkeley C.V. and Daulat Ramamoorthy Distinguished Research Award |
| 2017 | Berkeley EECS Rising Star |
| 2018 | MIT Technology Review Innovators Under 35 (TR35) |
| 2019 | ACM Doctoral Dissertation Award |
| 2020 | Microsoft Research Faculty Fellowship |
| 2020 | Samsung AI Researcher of the Year |
| 2020 | Intel Rising Star Faculty Award |
| 2021 | Office of Naval Research Young Investigator Award |
| 2022 | IEEE Robotics and Automation Society Early Academic Career Award |
| 2023 | Alfred P. Sloan Research Fellowship (Sloan Research Fellowship) |
| 2025 | Presidential Early Career Award for Scientists and Engineers (PECASE) |
The PECASE, which Finn received together with Stanford colleagues Dorsa Sadigh and Mary Wootters and which was announced in early 2025, is described by the White House as "the highest honor bestowed by the United States Government on outstanding scientists and engineers early in their careers." [9] In addition, Finn's coauthored papers have received best paper or outstanding paper recognition at venues such as the Conference on Robot Learning, the IEEE International Conference on Robotics and Automation, and the International Conference on Machine Learning.
Finn is an active public communicator on topics in robotics and machine learning. She has appeared in mainstream coverage in The New York Times, Wired (including its "5 Levels of Difficulty" video on machine learning), Bloomberg, MIT Technology Review, and TechCrunch, and her group's robot demonstrations are regularly featured in technology and science press. [14] She has given invited talks and keynote addresses at NeurIPS workshops, the Conference on Robot Learning, the AAAI Conference on Artificial Intelligence, the Robotics: Science and Systems conference, IROS, and ICRA. She has also delivered seminars at universities including MIT, Stanford, UC Berkeley, Carnegie Mellon, and Princeton.
Finn has co-organized multiple meta-learning workshops at NeurIPS and ICML, the BAIR Open Research Commons workshops at Berkeley, and is regularly a member of program committees and area chair lists at top-tier conferences in machine learning and robotics. She has supported diversity initiatives including Women in Machine Learning (WiML), Berkeley WiCSE, and the EECS Rising Stars workshop, and has helped develop AI outreach activities for K-12 and undergraduate students.