ABAB Pedia

Jost Tobias Springenberg

Jost Tobias Springenberg, known as Tobi, is a machine learning and robotics researcher at Physical IntelligencePhysical IntelligencePhysical Intelligence (π or Pi for short) is a San Francisco–based robotics AI company that develops general-purpose foundation models for many kinds of robots; its best-known work includes the π0 family of models.Open the full entry. Born in Hilden, Germany, he studied cognitive science and computer science and worked at the University of Freiburg and Google DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry. His projects include an all-convolutional vision network, Gato, RoboCat and the π0.5 model for household robots.

Contents6 sections
Key facts

Early life and education: From cognitive science to computer vision

Springenberg was born in Hilden, Germany, and goes by Tobi. He earned a bachelor’s degree in cognitive science at the University of Osnabrück, followed by a master’s in computer science at the University of Freiburg, where he worked with the machine learning and computer vision groups. He wrote that his undergraduate studies left him with a lasting interest in human cognition. He also publishes photographs on his personal website.⁠[1][2]

Testing what an image recognition network needs

In December 2014 he published work with Alexey DosovitskiyAlexey DosovitskiyAlexey Dosovitskiy is a computer vision and machine learning researcher and a Recursive co-founder, with contributions to FlowNet, CARLA and Vision Transformer.Open the full entry, Thomas Brox and Martin Riedmiller testing whether image recognition networks needed their customary max-pooling stages. On several benchmarks the team could use convolutional layers throughout and retain competitive performance. The paper was accepted to an ICLR workshop in 2015. That year he also collaborated with Tobias Domhan and Frank Hutter on using early learning curves to estimate later performance and speed up the selection of neural network settings.⁠[3][4]

Through 2024: Studying how robots learn at Google DeepMind

In Google DeepMind’s control team he studied applications of machine learning to science and robotics. His personal biography from that period listed London as his home. He contributed to Gato, helping deploy a model capable of varied tasks on a physical robot. He also worked on RoboCat, which explored how a robot could gain skills across different manipulation tasks and hardware.⁠[1][5]

He was first author of a study presented at the 2024 International Conference on Machine Learning. The team brought together 132 continuous-control tasks in one dataset, combining expert demonstrations with less successful attempts so they could test whether a model could learn multiple skills from varied past experience.⁠[6]

Household robotics and precise manipulation at Physical Intelligence

In April 2025 Physical Intelligence named Springenberg among the authors of π0.5. The team evaluated robots in homes absent from the training data, asking them to tidy kitchens and put away belongings in bedrooms to test how well they worked with unfamiliar objects and arrangements. He also represented the company at the AI Engineer conference that year.⁠[7][5]

In a January 6, 2026 interview with Karol HausmanKarol HausmanKarol Hausman is a robotics and artificial intelligence researcher who co-founded Physical Intelligence, a robot foundation-model company, and serves as its chief executive. He previously worked on robot learning at Google Brain and Google DeepMind and is an adjunct professor at Stanford University.Open the full entry, Springenberg discussed the distance between an occasional success and a robot reliable enough to use every day. As examples of progress, he cited roughly doubled box-assembly throughput and demonstrations lasting about 13 hours for espresso and four hours for laundry. On March 19 he and colleagues including Charles Xu published work on precise robot manipulation, measuring the most demanding stages of tasks such as driving a screw, fastening a zip tie and plugging in cables.⁠[8][9][10]

Related Organizations, Websites and Public Accounts

University of Freiburg: Former researcher in the machine learning and computer vision groups. Official website: https://uni-freiburg.de/ (opens in a new window).⁠[1][4]

Google DeepMind: Former research scientist in the control team. Official website: https://deepmind.google/ (opens in a new window).⁠[1][5]

Physical Intelligence: Robotics researcher. Official website: https://www.pi.website/ (opens in a new window).⁠[10][5]

Personal website: https://springenberg.aldeia.net/ (opens in a new window). He uses stokasto on GitHub and @jtspringenberg on X. His Google Scholar profile lists his research publications. https://scholar.google.com/citations?user=MGXJkIAAAAAJ (opens in a new window).⁠[1][11][5]

Sources11 sources

  1. Jost Tobias Springenberg (opens in a new window)springenberg.aldeia.net↑
  2. Photography (opens in a new window)springenberg.aldeia.net↑
  3. Striving for Simplicity: The All Convolutional Net (opens in a new window)arxiv.org↑
  4. Jost Tobias Springenberg — IJCAI 2015 (opens in a new window)ijcai.org↑
  5. Jost Tobias Springenberg: Bio, Work & Ideas (opens in a new window)ai.engineer↑
  6. Offline Actor-Critic Reinforcement Learning Scales to Large Models (opens in a new window)proceedings.mlr.press↑
  7. A VLA with Open-World Generalization (opens in a new window)pi.website↑
  8. Training General Robots for Any Task: Physical Intelligence's Karol Hausman and Tobi Springenberg (opens in a new window)sequoiacap.com↑
  9. Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg (opens in a new window)youtube.com↑
  10. Precise Manipulation with Efficient Online RL (opens in a new window)pi.website↑
  11. stokasto (Jost Tobias Springenberg) (opens in a new window)github.com↑