Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg

Summary

Karol Hausman and Tobi Springenberg of Physical Intelligence discuss why robotics has been bottlenecked by intelligence rather than hardware. They explain their end-to-end vision-language-action foundation models, including π*0.6, which uses real-world RL from experience to improve reliability and deployment. They highlight progress in generalization, data collection, value functions, and applications from coffee-making to laundry, while noting commercialization remains early.

• Physical Intelligence builds robotic foundation models intended to let any robot do any task.
• The founders argue intelligence, not hardware, is the main robotics bottleneck.
• End-to-end VLA models combine internet-scale pretraining with robot action data.
• π*0.6 adds RL from real-world experience, improving throughput and failure recovery.
• Real-world deployment is prioritized over simulation for manipulation due to long-tail failures.
• Data quality, diversity, and autonomous deployment data are seen as key scaling factors.
• Generalization across tasks and embodiments remains an open but improving challenge.
• Commercialization model and customer deployment strategy are still being determined.