Training Data Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg
147 snips
Jan 6, 2026 
Karol Hausman and Tobi Springenberg from Physical Intelligence discuss the groundbreaking potential of robotic foundation models. They argue that the intelligence bottleneck, not hardware, limits robotics and explain their mission to create models capable of performing diverse tasks. The duo dives into their end-to-end learning approach, emphasizing recent improvements in reinforcement learning and real-world deployment. Insights into unexpected applications from open-sourced models and the aspiration for continual robot learning highlight a pivotal shift in intelligent machine design.
AI Snips
Chapters
Transcript
Episode notes
Teleoperation Shows Hardware Is Capable
- Karol recounts teleoperated robots that could clean houses if a human controlled them, proving hardware can already do many tasks.
- This example motivated focusing on intelligence as the core bottleneck for robotics progress.
Three Axes For Scaling Robots
- They judge progress by capability, generalization, and performance as separate scaling challenges.
- Diversity of data is the primary path to generalization; deployment unlocks massive data collection.
Vision-Language-Action Architecture
- Their architecture extends vision-language models with an action expert to map images and instructions to robot actions.
- They pre-train mostly on their own robotics data mixed with internet data and use large transformers as backbones.