Karol Hausman and Tobi Springenberg of Physical Intelligence discuss how robotic foundation models can learn tasks, generalize across settings, and improve through real-world experience.
Hausman and Springenberg argue that robot hardware can already perform many tasks when guided by a person; the bigger limitation is the intelligence needed to operate independently. Physical Intelligence focuses on that problem, aiming to support many kinds of robots and applications rather than building a robot for just one job.
Hausman traces robotics’ shift from hand-written rules to learned perception, planning, and control components, then to end-to-end models that map sensory input to actions. Physical Intelligence builds on vision-language models with an action component, using internet pretraining alongside robot data and human demonstrations. The team sees diverse data as central to generalization, though it says the field still lacks a rigorous way to measure data quality and diversity.
For Pi-Star 0.6, the team starts with a policy trained on demonstrations, then improves it using robot attempts, human success ratings, and corrections. The work uses real robots because manipulation can expose failures that are hard to capture in simulation, such as cardboard sheets sticking together. Across box building, espresso preparation, and laundry folding, policy throughput more than doubled.
Hausman says robots ran for hours, including making coffee for 13 hours and folding laundry for four, while recovering from more failures than the demonstration-trained models. He describes reliability over long runs as a key step toward practical use. Feedback can also fix specific behaviors: 30 to 50 human corrections helped a coffee-making robot tamp grounds more gently.
Springenberg says corrections may initially improve the task they address, while transfer between tasks could grow when they share motions; broader pretraining also helps new tasks become easier to add. The team expects deployed robots to return useful experience for future training, but says its commercial model is still undecided. Hausman notes that general-purpose robotics may face a different reliability bar from self-driving, since some everyday tasks can tolerate occasional imperfect results.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Karol Hausman, co-founder and CEO, and Kevin Black, a researcher at Physical Intelligence, discuss robot foundation models and the path toward models that can control many kinds of robots with Going Direct’s host.
