Karol Hausman

Google DeepMind to Physical Intelligence Co-founder

Karol Hausman

Home > People > Karol Hausman

Profile

FieldDetails
Current PositionPhysical Intelligence Co-founder
PreviousGoogle DeepMind Staff Research Scientist
PhDUSC (University of Southern California)
NationalityPolish

Key Contributions

  • RT Series Key Leader: Led development of RT-1, RT-2, RT-X
  • SayCan: Early research connecting LLMs to robot control
  • Physical Intelligence Founding: pi0 development
  • Google Robotics Key Figure: Led industrialization of VLA research

Research Timeline

PhD & Early Career (2012-2017)

USC - Advised by Stefan Schaal

YearWorkImpact
2015Skill LearningRobot skill learning
2017Multi-Task LearningMulti-task robot learning

Google Brain / DeepMind (2017-2024)

Core Google Robotics Research

YearWorkImpact
2018JoinedGoogle Brain Robotics
2020Multi-Task RLMulti-task learning
2022SayCanLLM + robot grounding
2022RT-1Robotics Transformer
2023RT-2First VLA model
2023RT-XOpen X-Embodiment
2023PaLM-EEmbodied Language Model

Physical Intelligence (2024-present)

Co-founding & pi0 Development

YearWorkImpact
2024Physical Intelligence FoundedGeneral-purpose robot AI
2024pi0Flow matching VLA
2025pi0.5Open-world generalization

Major Publications

VLA & Foundation Models

  • RT-1 (2022) - Robotics Transformer
  • RT-2 (2023) - Vision-Language-Action
  • RT-X (2023) - Open X-Embodiment
  • PaLM-E (2023) - Embodied multimodal model
  • pi0 (2024) - Flow matching VLA

LLM + Robotics

  • SayCan (2022) - LLM grounding to robotics
  • Inner Monologue (2022) - LLM feedback for robots

Multi-Task Learning

  • Multi-Task RL (2018)
  • Skill Composition (2019)

Key Ideas

SayCan (2022)

Core: Combining LLM language understanding with robot's actual execution capability

Feasibility = P(useful|LLM) x P(success|Robot)

LLM: "What should be done" (semantic)
Robot: "What can be done" (affordance)

Impact:

  • Core early research in LLM + robot integration
  • Foundation for many subsequent LLM-robot studies

RT-2 & VLA (2023)

Core: Using VLM directly for robot control

Previous: Separate perception + planning + control
RT-2: Single VLM directly outputs image-to-action

Impact:

  • Established VLA paradigm
  • Foundation model application to robotics

Philosophy & Direction

Research Philosophy

“The key to robot AI is generalization. The goal is general capability, not specific tasks.”

Research Direction Evolution

  1. 2012-2017: Skill learning, multi-task RL
  2. 2017-2022: Large-scale robot learning at Google
  3. 2022-2023: LLM + robotics, VLA models
  4. 2024-present: Foundation models, Physical Intelligence

Google to Physical Intelligence

Achievements at Google

  • Established VLA paradigm with RT series
  • Initiated LLM-robot integration with SayCan
  • Key figure in Google Robotics

Motivation for Founding

  • Beyond academic research to actual productization
  • Commercializing general-purpose robot AI
  • Fast execution and focus

See Also