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

Physical Intelligence’s Hausman & Springenberg: why robotics was always an intelligence bottleneck, and how π0.6’s real-world RL just crossed the deployment threshold*

• Intelligence, not hardware, is the bottleneck — teleoperated robots could clean entire houses 10+ years ago; hardware capability was never the limit.
• Classical perception→planning→control pipeline was fundamentally wrong; interfaces between components are what break, not the components themselves.
• π*0.6 uses RL from real-world experience (not sim) — 2x+ throughput on 3 tasks: espresso making, cardboard box folding, laundry; robot ran coffee for 13 hours straight.
• Only 30–50 human correction episodes reshaped tamping force behavior in a model pre-trained on millions of episodes — scale of prior training makes small corrections surprisingly effective.
• Sim-to-real works for locomotion (model your own body once) but fails for manipulation (you’d have to simulate every object in the world).
• Value functions trained on the same backbone predict failure 30–40 steps before it occurs — enabling early data filtering without waiting for terminal reward.
• Deployment data will eventually dwarf internet + teleoperation data combined; the bootstrap phase is just to reach the deployment threshold, after which data collection becomes self-funding.
• π0 models generalize across surgical robots, agricultural robots, drones, and driving from the same weights — cross-domain transfer mechanisms are not yet understood.
• PI deliberately refuses to pick a vertical application, citing history of robotics startups that narrowed into ‘warehouse pick-and-place’ companies and got stuck.
• Thought commercial deployment was 5 years out; crossed the threshold ~2 months before recording (≈ Nov 2025) ahead of schedule.

Guests: Karol Hausman (Physical Intelligence co-founder), Tobi Springenberg (Physical Intelligence co-founder) · 2026-01-06 · Watch on YouTube