Chelsea Finn

ref · May 25, 2026, 6:26am

Assistant Professor of Computer Science and Electrical Engineering at Stanford University, inventor of Model-Agnostic Meta-Learning (MAML), co-author of Direct Preference Optimization (DPO), and co-founder of Physical Intelligence — one of the most cited early-career researchers in machine learning and robot learning.

Profile

Field	Detail
Born	October 8, 1992, United States
Nationality	American
Current Institutions	Stanford University (IRIS Lab); Physical Intelligence (Co-founder)
Research Areas	Meta-Learning, Robot Learning, Reinforcement Learning, Imitation Learning, Vision-Language-Action Models
PhD Advisors	Sergey Levine; Pieter Abbeel
PhD Dissertation	Learning to Learn with Gradients (UC Berkeley, 2018)
Academic Website	ai.stanford.edu/~cbfinn
X / Twitter	@chelseabfinn
GitHub	@cbfinn
Google Scholar	scholar.google.com (126,000+ citations)

Overview

Chelsea Finn is an American computer scientist and one of the most influential researchers of her generation at the intersection of machine learning and robotics. She is best known for inventing Model-Agnostic Meta-Learning (MAML), a gradient-based algorithm published at ICML 2017 that became one of the most cited papers of the decade and established meta-learning as a major research direction. As a co-author of Direct Preference Optimization (DPO, NeurIPS 2023), she contributed a second highly impactful technique, this time for language model alignment, which rapidly displaced RLHF-based approaches in both academic and industrial settings. Finn directs the IRIS Lab (Intelligence through Robotic Interaction at Scale) at Stanford, where she holds joint appointments in Computer Science and Electrical Engineering as the William George and Ida Mary Hoover Faculty Fellow. In 2024 she co-founded Physical Intelligence (Pi), a robotics AI company that raised $400 million at a $2.8 billion valuation and is developing general-purpose foundation models for physical robots. She has accumulated over 126,000 Google Scholar citations — an exceptional figure for a researcher who completed her PhD in 2018.

Early Life & Education

B.S., Electrical Engineering and Computer Science — Massachusetts Institute of Technology
Finn completed her undergraduate degree at MIT in EECS, building the technical foundation in machine learning and control systems that would frame her graduate research.

Ph.D., Computer Science — University of California, Berkeley, 2018
Finn joined the Berkeley Artificial Intelligence Research Lab (BAIR), advised jointly by Sergey Levine and Pieter Abbeel. Her dissertation, Learning to Learn with Gradients, developed the theoretical and algorithmic framework for gradient-based meta-learning — the idea that a model’s parameters can be optimized to serve as an excellent initialization for rapid adaptation to new tasks via a small number of gradient steps. The dissertation received the ACM Doctoral Dissertation Award, the field’s highest recognition for a PhD thesis. She was the first woman to receive the UC Berkeley EECS C.V. & Daulat Ramamoorthy Distinguished Research Award (2016), given for exceptional research during doctoral study. During her PhD she also interned at Google Brain, working on large-scale robot learning from deep predictive models.

Career

UC Berkeley — BAIR Lab (2014–2018)

As a PhD student under Levine and Abbeel, Finn developed the methods that would define her early reputation: end-to-end learning of visuomotor policies, video prediction for physical interaction, and — most significantly — the MAML algorithm. Alongside her doctoral research she co-taught a pioneering course on deep reinforcement learning (UC Berkeley CS294-112, Spring 2017), co-organized a workshop on deep learning for action and interaction at NeurIPS 2016, and co-organized BAIR Camp, an outreach summer camp for high school students from low-income backgrounds.

Google Brain — Internship (2016–2017)

During her PhD, Finn worked as a research intern at Google Brain, contributing to robot learning projects involving large-scale predictive model training and robotic grasping datasets, several of which were released publicly as the Google Brain Robotics Data corpus.

Stanford University — IRIS Lab (2019–present)

Finn joined Stanford as an assistant professor in 2019. She holds joint appointments in CS and EE and is affiliated with Stanford HAI and the Stanford ML Group. Her lab, IRIS (Intelligence through Robotic Interaction at Scale), pursues research across four themes: learning from offline data and demonstrations, meta-learning and generalization, vision-language-action models, and real-world robot deployment.

CS330: Deep Multi-task and Meta Learning — Finn designed and taught this Stanford course beginning in Fall 2019; lecture videos from multiple editions are publicly available and have been widely adopted as a standard curriculum for graduate-level meta-learning study globally.

CS224R: Deep Reinforcement Learning — Finn also teaches this Stanford course, covering modern deep RL methods and their application to robotic control.

Key research threads at Stanford:

During her faculty tenure, Finn’s lab has contributed to several distinct research streams. In meta-learning and generalization, work on out-of-distribution robustness and distribution shift (including the WILDS benchmark, ICML 2021) extended her PhD-era contributions to the broader problem of robust ML. In robot learning, her group has worked on offline and online RL for robotics, cross-embodiment transfer (Bridge Data series), and humanoid imitation (HumanPlus, 2024). Most recently, OpenVLA (2024) — an open-source vision-language-action model co-developed with collaborators at Stanford and Berkeley — set a new open standard for generalist robot control. In language models, DPO (2023) became one of the most consequential contributions to LLM alignment methodology in recent years.

Physical Intelligence (Pi) — Co-Founder (2024–present)

In 2024, Finn co-founded Physical Intelligence alongside Sergey Levine, Karol Hausman, Brian Ichter, Lachy Groom, Adnan Esmail, and Quan Vuong. The company’s thesis — that a single large foundation model can learn to control any robot for any task, analogously to how large language models generalized across text — directly reflects the research program Finn and Levine developed over a decade. Physical Intelligence has raised $400 million at a valuation of $2.8 billion, with backing from Bezos Expeditions, Khosla Ventures, and the OpenAI Startup Fund. The company’s first public model, π0 (pi-zero), is a vision-language-action model demonstrating generalist dexterous manipulation across multiple robot platforms.

Key Contributions

Model-Agnostic Meta-Learning (MAML) — Presented at ICML 2017 (Finn, Abbeel, Levine), MAML introduced a gradient-based meta-learning algorithm that optimizes initial model parameters for rapid adaptation to new tasks with minimal data. Its model-agnostic design meant it applied to any gradient-trained model — classification, regression, or reinforcement learning — and it became one of the foundational papers of the meta-learning literature, accumulating tens of thousands of citations and spawning dozens of algorithmic variants including Reptile, FOMAML, and iMAML.
End-to-End Deep Visuomotor Policies — Published in JMLR 2016 (Levine, Finn, Darrell, Abbeel), this paper demonstrated that deep neural networks could learn robot manipulation policies end-to-end from camera images and joint angles, bridging visual perception and motor control in a single differentiable system.
Direct Preference Optimization (DPO) — Published at NeurIPS 2023 (Rafailov, Sharma, Mitchell, Ermon, Manning, Finn), DPO recast the RLHF preference learning objective as a supervised binary classification problem, eliminating the need for a separate reward model and the instabilities of PPO-based fine-tuning. It was almost immediately adopted across both the research community and industry, and has accumulated citations at one of the fastest rates in recent NeurIPS history. Finn served as a senior co-author providing the lab infrastructure and research direction.
CS330: Deep Multi-task and Meta Learning — This Stanford graduate course created by Finn in 2019 became the primary university-level curriculum for meta-learning and multi-task learning globally; its publicly posted video lectures have reached practitioners and researchers far beyond Stanford’s enrollment.
OpenVLA: An Open-Source Vision-Language-Action Model — Published at CoRL 2024 (Kim, Pertsch, Karamcheti, et al., co-supervised by Finn and collaborators), OpenVLA established a freely available, reproducible baseline for generalist robot control that democratized access to VLA research for the broader community.
Open X-Embodiment and RT-X — Finn was part of the large collaborative team (ICRA 2024) that assembled the Open X-Embodiment dataset spanning 22 robot embodiments and 527 skills, enabling cross-embodiment transfer learning and releasing it publicly; a landmark step toward generalizable robot foundation models.
HumanPlus: Humanoid Shadowing and Imitation — Published at CoRL 2024 (Fu, Zhao, Wu, Wetzstein, Finn), this work demonstrated learning of whole-body humanoid manipulation from human motion capture via a combination of imitation learning and reinforcement learning.
WILDS Benchmark — Co-authored (Koh et al., ICML 2021), WILDS provided a rigorous benchmark for distribution shift in machine learning across ten real-world datasets, establishing a standard for evaluating robustness to deployment shifts.
Video Prediction for Physical Interaction — Published at NeurIPS 2016 (Finn, Goodfellow, Levine), this work demonstrated that neural networks could learn forward models of physical scenes from raw video, enabling unsupervised learning of robot interaction dynamics.

Awards & Recognition

Presidential Early Career Award for Scientists and Engineers (PECASE) (2025) — The United States government’s highest honor for early-career scientists and engineers.
Alfred P. Sloan Research Fellowship (2023) — Awarded to outstanding early-career faculty.
IEEE RAS Early Academic Career Award in Robotics and Automation (2022)
Office of Naval Research Young Investigator Award (2021)
Microsoft Research Faculty Fellowship (2020)
Samsung AI Researcher of the Year (2020) — Awarded to 19 early-career AI researchers globally.
Intel Rising Star Faculty Award (2020) — Awarded to 7 US faculty in information and telecommunications.
ACM Doctoral Dissertation Award (2019) — The ACM’s highest recognition for a doctoral thesis in computing; awarded for Learning to Learn with Gradients.
MIT Technology Review Innovators Under 35 — Pioneer (2018)
C.V. & Daulat Ramamoorthy Distinguished Research Award, UC Berkeley EECS (2016) — First woman to receive this award.

Key Relationships

Sergey Levine — PhD co-advisor at Berkeley; Physical Intelligence co-founder; Finn and Levine’s decade-long collaboration spans visuomotor policy learning, offline RL, cross-embodiment transfer, and the founding of Pi; among the most productive research partnerships in robot learning.
Pieter Abbeel — PhD co-advisor at Berkeley; MAML was co-authored with Abbeel; his lab’s emphasis on meta-learning and robotic manipulation shaped Finn’s foundational contributions; also a co-author on early visuomotor policy work.
Rafael Rafailov — PhD student at Stanford whose dissertation work Finn co-supervised; first author of DPO, one of the most impactful alignment papers of 2023.
Fei-Fei Li — Senior Stanford colleague and faculty affiliate of HAI; appears among Finn’s most prominent academic followers and professional network; their Stanford lab affiliations overlap in robotic perception and learning.
Karol Hausman — Physical Intelligence co-founder; former Google Brain researcher whose work on RT-2 and large-scale robot learning directly converged with Finn’s agenda.
Tony Z. Zhao — PhD student and key contributor to the ALOHA/Unleashed and mobile manipulation work coming out of Finn’s broader Stanford network.
Ian Goodfellow — Co-author on the 2016 NeurIPS video prediction paper; an early collaborator who appears among Finn’s most prominent professional connections.
Christopher Manning — Stanford colleague and co-author on DPO; the collaboration exemplifies Finn’s cross-disciplinary engagement between robotics and NLP at Stanford.

Personal Style

Finn’s research is characterized by a strong preference for simple, general algorithms over those engineered for specific domains — a sensibility that directly produced MAML’s model-agnostic design and DPO’s reformulation of RLHF as supervised classification. Her lab’s framing of “intelligence through robotic interaction at scale” reflects a conviction that embodied interaction with the physical world is both the key test and the key driver of general intelligence. In course design and outreach, she has invested substantially in making graduate-level research methods widely accessible — both through publicly posted Stanford lecture series and through diversity-focused education programs at Berkeley. Her public communication is technically precise and notably low on hype; she tends to frame new results in terms of their failure modes and remaining open problems rather than superlatives.