Training General Robots for Any Task: Physical Intelligence’s Karol Hausman and Tobi Springenberg
Physical Intelligence’s Karol Hausman and Tobi Springenberg argue robotics is bottlenecked by intelligence, not hardware, and that π0.6 proves RL-from-experience can make robots deployable today.*
- π*0.6 increased task throughput by over 2x on three real-world tasks: box-building, espresso-making, and laundry-folding.
- A robot ran continuously for 13 hours making coffee and 4 hours folding laundry — sustained deployment, not cherry-picked demos.
- Physical Intelligence crossed the commercial deployment threshold roughly two months before this recording, two to three years ahead of their own internal estimate.
- RL is done entirely on real robots, not simulation, because manipulation failures (e.g., imperfectly perforated cardboard sticking together) don’t appear in sim.
- Just 30–50 human corrections on tamping force were enough to retrain a model pre-trained on millions of episodes — pointing toward viable continual learning.
- The model generalizes across radically different embodiments — surgical robots, drones, espresso machines — via a shared vision-language-action backbone; the team says they don’t fully understand why this works.
- Hausman’s thesis: once robots are widely deployed, autonomous experience data will dwarf internet data as the primary training source, making deployment data cost-negative.
- Classical pipeline robotics (separate perception, planning, control modules) failed because the interfaces between modules, not the modules themselves, were the breaking point.
2026-01-06 · Watch on YouTube
