# Anca Dragan

> Source: https://aiwiki.ai/wiki/anca_dragan
> Updated: 2026-06-07
> Categories: People
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

Anca Diana Dragan is a Romanian-American computer scientist and roboticist known for her work on human-robot interaction and [AI alignment](/wiki/ai_alignment). She is an associate professor of electrical engineering and computer sciences at the [University of California, Berkeley](/wiki/uc_berkeley), where she founded and directs the InterACT Lab, and since 2024 she has led the [AI safety](/wiki/ai_safety) and alignment organization at [Google DeepMind](/wiki/google_deepmind). In that role she oversees the teams responsible for the safety and alignment of the [Gemini](/wiki/gemini) family of models, and she also co-leads Gemini post-training.[1][2]

Dragan's academic research helped establish a mathematical foundation for robots and other AI agents that act with and around people, including influential work on legible robot motion and on learning objectives from human feedback. She is a co-principal investigator of Berkeley's [Center for Human-Compatible AI](/wiki/center_for_human_compatible_ai) and a steering committee member of the [Berkeley Artificial Intelligence Research](/wiki/berkeley_ai_research) (BAIR) Lab.[1][3]

## Early life and education

Dragan was born in Braila, Romania.[4] She studied computer science at Jacobs University Bremen in Germany (now Constructor University), receiving her Bachelor of Science degree in 2009.[3][4] She then moved to the United States for doctoral study at the Robotics Institute of [Carnegie Mellon University](/wiki/carnegie_mellon_university) in Pittsburgh, where she worked in the Personal Robotics Lab with adviser [Siddhartha Srinivasa](/wiki/siddhartha_srinivasa). She completed her PhD in robotics in 2015.[1][3]

Her doctoral research centered on the difference between motion that is merely predictable and motion that is legible, meaning motion that actively communicates a robot's intent to a human observer. The 2013 paper "Generating Legible Motion," written with Srinivasa, introduced a formal optimization framework distinguishing the two properties and became one of her most cited early contributions.[5]

## Career

Dragan joined UC Berkeley as an assistant professor in the Department of Electrical Engineering and Computer Sciences in 2015, immediately after finishing her PhD, and was later promoted to associate professor with tenure.[1] At Berkeley she founded the InterACT Lab, whose research sits at the intersection of robotics, machine learning, and human-computer interaction. The lab's stated aim is to enable AI agents, ranging from robot arms to cars to large language models and recommender systems, to work with, around, and in support of people.[1]

In 2016 she became a co-principal investigator of the newly established Center for Human-Compatible AI (CHAI), founded by [Stuart Russell](/wiki/stuart_russell) at Berkeley, alongside fellow co-PIs including [Pieter Abbeel](/wiki/pieter_abbeel) and Tom Griffiths.[6] She has also been a co-founder and steering committee member of the BAIR Lab.[1]

Beyond academia, Dragan spent roughly six years working part-time as a senior research scientist at [Waymo](/wiki/waymo), Alphabet's autonomous-driving company, advising on the deployment of learning-based, safety-critical systems. Many of her former PhD students have gone on to faculty positions at institutions such as MIT, Stanford, Carnegie Mellon, and Princeton, and to research roles in industry.[1]

### Move to Google DeepMind

On March 28, 2024, Berkeley's EECS department announced that Dragan had been named Head of AI Safety and Alignment at Google DeepMind, a new organization the lab had formed in February 2024 to develop safeguards for its Gemini models and to align future models with human goals and values.[2] Dragan took leave from her Berkeley professorship to take up the role while remaining affiliated with the university.[1][2]

At DeepMind she leads a collection of teams covering both present-day model safety and longer-horizon alignment work in anticipation of more capable systems, including [artificial general intelligence](/wiki/artificial_general_intelligence). The organization's scope spans alignment research areas such as mechanistic interpretability and scalable oversight, as well as a Frontier Safety group that develops dangerous-capability evaluations and maintains DeepMind's Frontier Safety Framework.[7] According to DeepMind's safety researchers, the leadership of the organization has included Dragan together with [Rohin Shah](/wiki/rohin_shah), Allan Dafoe, and Dave Orr, with [Shane Legg](/wiki/shane_legg) as executive sponsor.[7] Dragan has also described herself as a co-lead of post-training for Gemini.[8]

As of 2026, her LinkedIn profile lists her title as Vice President of AI Safety, Alignment, and Collaboration at Google DeepMind, indicating an expansion of her original remit.[9] She remains on leave from UC Berkeley.[8]

## Research contributions

Dragan's research program has consistently treated the question of what an AI system should optimize for as central, an emphasis she summarized in her 2025 [International Conference on Machine Learning](/wiki/icml) invited talk, titled "What to optimize for, from robot arms to frontier AI." Her argument there was that deciding what to optimize for can matter more than how well a system solves the optimization problem.[8]

Her work spans several connected threads:

| Theme | Idea | Notable work |
| --- | --- | --- |
| Legible motion | Robots should move in ways that communicate intent, not merely reach goals predictably | "Generating Legible Motion," RSS 2013[5] |
| Cooperative reward learning | Treat value alignment as a cooperative game in which a robot infers a human's objective | Cooperative Inverse Reinforcement Learning, NeurIPS 2016[6] |
| Learning from diverse feedback | Infer objectives from comparisons, physical corrections, and demonstrations, not just labels | InterACT Lab reward-learning program[1] |

A landmark result is Cooperative Inverse Reinforcement Learning (CIRL), introduced in a 2016 paper with [Dylan Hadfield-Menell](/wiki/dylan_hadfield_menell), Pieter Abbeel, and Stuart Russell. CIRL frames alignment as a two-player game of partial information in which a robot does not know the true reward function and must learn it from a human, who in turn may behave differently when teaching than when simply performing a task. The formulation became a widely referenced model for thinking about how an agent uncertain about human objectives has an incentive to seek and defer to human oversight.[6] Dragan and Russell later extended the line of work with results such as an efficient, generalized Bellman update for CIRL.[6]

A unifying theme across these efforts is that human behavior is itself informative data about human goals. Dragan's lab has argued that an AI system can recover what people actually want by combining many channels of evidence, including stated preferences, corrections to a robot's behavior, and even the reward function a designer writes down, all treated as noisy observations of an underlying objective rather than as ground truth.[1] This research framing carried directly into her DeepMind work on aligning large models with human values while accounting for the plurality and diversity of those values.[2]

## Recognition

Dragan has received numerous early-career honors recognizing both her research and her teaching. Her work has also collected best-paper recognition at major robotics venues including RSS, ICRA, IROS, and the ACM/IEEE International Conference on Human-Robot Interaction.[3]

| Year | Honor |
| --- | --- |
| 2017 | MIT Technology Review 35 Innovators Under 35[3] |
| 2017 | NSF CAREER Award[3] |
| 2017 | Okawa Foundation Research Grant[3] |
| 2018 | Sloan Research Fellowship[3] |
| 2019 | Presidential Early Career Award for Scientists and Engineers (PECASE)[3] |
| 2020 | Office of Naval Research Young Investigator Award[3] |
| 2020 | UC Berkeley McEntyre Award for Excellence in Teaching[3] |
| 2021 | IEEE RAS Early Academic Career Award in Robotics and Automation[3] |

She has served the research community in roles including program co-chair of the Conference on Robot Learning in 2018 and senior area chair for the International Conference on Learning Representations, and as an area chair or associate editor for venues such as ICRA, IROS, and the journals Transactions on Human-Robot Interaction and Autonomous Robots.[3] She has delivered keynote and invited talks at conferences including IROS, CoRL, ICAPS, and ICML.[3][8]

Dragan has become a visible public voice on AI safety since joining DeepMind, including an appearance discussing safety and existential risk on the official Google DeepMind podcast, and she has argued publicly that progress on safety requires close collaboration between industry and academia.[7]

## References

1. Anca Dragan, personal website, EECS, University of California, Berkeley. https://people.eecs.berkeley.edu/~anca/ (accessed 2026).
2. "Anca Dragan named Head of AI Safety and Alignment at Google DeepMind," EECS at Berkeley, March 28, 2024. https://eecs.berkeley.edu/news/anca-dragan-named-head-of-ai-safety-and-alignment-at-google-deepmind/
3. "Activities and Awards," Anca Dragan, EECS, UC Berkeley. https://people.eecs.berkeley.edu/~anca/activities-awards.html
4. "Anca Dragan," Constructor University (formerly Jacobs University Bremen) alumni gallery. https://constructor.university/more/alumni-office/alumni-gallery/anca_dragan
5. Anca Dragan and Siddhartha Srinivasa, "Generating Legible Motion," Robotics: Science and Systems IX, 2013.
6. Dylan Hadfield-Menell, Stuart J. Russell, Pieter Abbeel, and Anca Dragan, "Cooperative Inverse Reinforcement Learning," Advances in Neural Information Processing Systems (NeurIPS), 2016. https://arxiv.org/pdf/1606.03137
7. "AGI Safety and Alignment at Google DeepMind: A Summary of Recent Work," DeepMind Safety Research, Medium, 2024.
8. "What to optimize for, from robot arms to frontier AI," Anca Dragan, ICML 2025 invited talk. https://icml.cc/virtual/2025/invited-talk/39874
9. Anca Dragan, LinkedIn profile, "VP, AI Safety, Alignment, and Collaboration." https://www.linkedin.com/in/ancadianadragan/