Karol Hausman
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,692 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Last reviewed
May 31, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,692 words
Add missing citations, update stale details, or suggest a clearer explanation.
Karol Hausman is a roboticist and entrepreneur who is co-founder and chief executive officer of Physical Intelligence, a San Francisco startup that builds foundation models for robotics. The company, often styled "Pi" or with the Greek letter "π," develops vision-language-action models intended to control many kinds of robots across many tasks, and it released the π0 (pi-zero) and π0.5 models in 2024 and 2025. [1][2][6] Before founding the company, Hausman was a staff research scientist at Google DeepMind, where he worked on robot learning projects such as SayCan and the RT-1 and RT-2 models, and he has served as an adjunct professor at Stanford University. [3][4] He holds a Ph.D. in robotics from the University of Southern California. [5][13]
Hausman studied mechatronics at the Warsaw University of Technology in Poland, where he earned a B.Sc. and an M.Sc. [3][13] He went on to a second master's degree in robotics at the Technical University of Munich in Germany, completing a thesis on object segmentation and recognition using interactive perception. [3][13]
He then moved to the United States for doctoral study at the University of Southern California. There he worked in the Robotic Embedded Systems Laboratory under the supervision of Gaurav Sukhatme, and he completed his Ph.D. in 2018. [5][13] His doctoral research dealt with robotic perception and control, including interactive and active perception, where a robot uses its own physical interaction with objects to perceive them better. [3][5] During this period he also collaborated with researchers connected to Stefan Schaal's robot learning group at USC. [5]
After his doctorate, Hausman joined Google, working first in the Google Brain team and later in Google DeepMind after Google merged its research groups. As a staff research scientist he focused on robot learning, reinforcement learning, and the use of large models for robot decision making. [3][4] His stated research interest was enabling robots to acquire general-purpose skills in the real world, with a growing emphasis on building and using foundation models for robotic control. [3]
Hausman was among the authors of SayCan, a 2022 project also titled "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances." SayCan paired a large language model with a set of learned low-level robot skills so that a mobile manipulator could carry out long, abstract instructions given in natural language. The language model proposed useful steps, and learned value functions estimated which of those steps the robot could actually perform in its current situation. [4][7] According to Hausman's professional profile, the SayCan paper received a special innovation award at the 2022 Conference on Robot Learning. [3]
He also contributed to the Robotics Transformer line of work. RT-1, published in 2022, was a transformer model trained on a large set of real robot demonstrations to perform many manipulation tasks from camera images and text commands. [4][8] RT-2, released in 2023, extended this idea by co-training on web image and text data together with robot data, producing an early vision-language-action model that could transfer some web-scale knowledge into robot behavior. [4][9] This research connected to broader efforts at Google such as PaLM-E and the Open X-Embodiment dataset, a shared collection of robot data gathered across many laboratories. [9][8]
Alongside this applied work, Hausman published earlier research on multi-task and self-supervised robotic reinforcement learning, imitation learning, and skill discovery, studying how a single robot could learn many behaviors and reuse them. [3] His Google Scholar record lists thousands of citations across these areas. [3]
While at Google, Hausman held an appointment as an adjunct professor at Stanford University. [3][4] He co-taught CS 224R, the university's course on deep reinforcement learning, which covers methods for training agents to act through trial and error and reward. [3] His teaching and research overlapped with the work of Stanford faculty in robot learning, and two collaborators from that environment, Sergey Levine and Chelsea Finn, later became his co-founders at Physical Intelligence. [10][11]
Hausman co-founded Physical Intelligence in 2024 and became its chief executive officer. [10][11] The founding group included Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail, and other researchers and company builders drawn from Google, Stanford, and the technology industry. [10][11]
The company's stated aim is to bring general-purpose artificial intelligence into the physical world by building learning algorithms for a single model that can control any robot to do any task. [2] Hausman has framed the goal as a general robot foundation model rather than software tied to one machine. In one account of the company's plans he said, "What we're doing is not just a brain for any particular robot. It's a single generalist brain that can control any robot." [11] The underlying premise is that the scaling behavior seen in language models might also apply to robot control if a model is trained on enough varied data across many robot bodies and environments. [1]
Physical Intelligence released its first model, π0, on October 31, 2024, describing it as a generalist robot policy. [6][12] π0 is a vision-language-action model: it takes in camera images and a text instruction and outputs continuous motor actions for a robot. [6][12] The model is built on PaliGemma, a vision-language model from Google with about three billion parameters, and it adds an "action expert" component that produces smooth, high-frequency motor commands using a technique related to flow matching. [12] π0 was trained on data from several different robots performing dozens of tasks, combined with the Open X-Embodiment dataset, and it could be fine-tuned to new tasks with roughly one to twenty hours of additional data. [6][12] In demonstrations it folded laundry, cleared and bussed tables, and handled other manipulation tasks. [6][12]
In February 2025 the company open-sourced π0, releasing the weights and code in an experimental repository called openpi, together with checkpoints for platforms such as the ALOHA arms and the DROID setup based on Franka arms. [12] In April 2025 it published π0.5, a successor described as a vision-language-action model with open-world generalization. [1] π0.5 was trained by co-training on heterogeneous data, including data from multiple robots, high-level semantic prediction, and web data, and it was shown controlling a mobile robot to clean kitchens and bedrooms in homes that did not appear in its training data, carrying out multi-step jobs lasting several minutes. [1] The company later announced further models in the same family. [2]
Physical Intelligence raised an early round of about 70 million dollars at a valuation near 400 million dollars, with backers including Thrive Capital, Khosla Ventures, Lux Capital, OpenAI, and Sequoia Capital. [11][14] In November 2024 it raised about 400 million dollars at a valuation around 2 billion dollars, in a round associated with Jeff Bezos, Lux Capital, and Thrive Capital, with participation from OpenAI, Redpoint Ventures, and Bond. [11][14] In November 2025 the company raised a further round reported at about 600 million dollars, valuing it near 5.6 billion dollars, with investors including CapitalG and Lux Capital, bringing total funding past one billion dollars. [15]
The work places Hausman within a wider research direction in embodied AI that tries to apply the methods of large pretrained models to physical robots. [1][2] Supporters of this approach argue that a single foundation model, trained on diverse robot data, could let robots generalize to new tasks and settings instead of being programmed for one job at a time. [1][11] The approach remains unproven at scale, and the π0 and π0.5 reports note that the models still make frequent mistakes in both planning and motor control. [1]
Hausman's research has been covered by general news outlets, including reports on Google's robot learning work and on Physical Intelligence's models. [3][6] His writing has also appeared in technical venues, and he has been featured in interviews about the company and the prospects for general-purpose robots. [1][11] He is listed as a co-founder of Physical Intelligence by several of the firm's investors. [10]
| Field | Detail |
|---|---|
| Name | Karol Hausman |
| Occupation | Roboticist, entrepreneur |
| Known for | Co-founder and CEO of Physical Intelligence; robot learning research (SayCan, RT-1, RT-2) |
| Current role | Co-founder and CEO, Physical Intelligence (since 2024) |
| Past roles | Staff research scientist, Google Brain and Google DeepMind; adjunct professor, Stanford University |
| Education | B.Sc. and M.Sc., mechatronics, Warsaw University of Technology; M.Sc., robotics, Technical University of Munich; Ph.D., robotics, University of Southern California (2018) |
| Doctoral advisor | Gaurav Sukhatme |
| Fields | Robotics, robot learning, reinforcement learning, vision-language-action models |
| Notable products | π0 (pi-zero), π0.5 robot foundation models |