Karol Hausman

Overview

Karol Hausman is a roboticist whose research covers robot learning, sim-to-real transfer and embodied AI ^[1]. The three terms describe one line of work. Instead of hand-coding every motion a machine must perform, researchers let it learn behaviour from data; much of that learning happens inside computer simulation, because real-world trials are slow and expensive to collect; and the aim is a system whose skills still hold up once the software is driving an actual physical machine. Embodied AI refers to exactly that setting, where perception, decision-making and action are coupled in a body that has to cope with the world rather than with a curated dataset.

Education

Hausman holds a PhD in computer science from the University of Southern California ^[1]. Before the doctorate he completed a master's degree in robotics at the Technical University of Munich ^[1]. He also holds master's and bachelor's degrees in mechatronics from Warsaw University of Technology ^[1]. Mechatronics is the field that joins mechanical engineering, electronics and control software, and it is a common starting point for people who later move into robotics research. The path from mechatronics through robotics to a computer science doctorate traces a wider shift in the field: from building machines to teaching them.

Research

His published work sits at the intersection of machine learning and control. A 2020 paper in the International Journal of Robotics Research proposed a skill latent space approach ^[1]. In general, a latent space is a compact set of learned variables that stands in for something large and messy — a collection of demonstration trajectories, for instance — so that a learning algorithm can work with the compressed version instead of the raw data. Representing skills this way is one route to reuse: a robot that has acquired a set of skills inside such a space can combine or adapt them, rather than learning each task from scratch.

A second thread is sim-to-real transfer, the problem of training a policy in simulation and then deploying it on hardware. Simulation offers unlimited practice and perfect labels, but it never matches the physical world exactly; the mismatch is usually called the reality gap. Work in this area tries to make training conditions varied enough, or the learned representation robust enough, that the gap does not break the policy once the robot leaves the simulator.

At Google, Hausman worked on learning large-scale generalisable grasping and manipulation policies ^[1]. That is the kind of system that has to pick up many different objects in many different arrangements, rather than repeat one scripted motion. Generalisation is the hard part: a policy tuned to one object or one table setup is of little use outside the lab.

Roles and service

He was previously a senior research scientist at Google Brain and Google DeepMind ^[1]. Since 2021 he has been an adjunct professor at Stanford University ^[1]. He has also served as an area chair or associate editor for ICRA, ICLR, NeurIPS and ICML ^[1]. Area chairs and associate editors run the review process for conferences and journals: they assign submissions to reviewers, read the resulting reports and help decide which papers are accepted. Serving in that capacity for both robotics venues such as ICRA and machine learning venues such as ICLR, NeurIPS and ICML reflects where his own work sits, between the two communities.

Physical Intelligence

Hausman co-founded Physical Intelligence and is listed on the company's team page ^[1][2]. For a researcher whose subject is generalisable manipulation, a company is one way to build such systems rather than write about them; the team page records him among the co-founders ^[2].

Significance

Taken together, his record combines academic teaching, industrial research and conference service. The research agenda — robot learning, transfer from simulation, embodied systems — is one that only works when the software meets real hardware, and his career runs through the institutions where that meeting is arranged.