Sergey Levine on Why Real-World Data Will Define Physical AI

Robotics foundation models require a transition from isolated laboratory science to large-scale industrial undertakings to achieve true generalization. Real-world data collection serves as the critical engine for this evolution, as robots deployed in diverse, dynamic environments generate the necessary experience to handle complex, long-horizon tasks. By integrating human feedback and reasoning-based supervision, these systems learn from failures and edge cases, effectively bootstrapping their own capabilities. This approach moves beyond specialized, repetitive automation toward a platform-agnostic intelligence capable of adapting to various hardware morphologies. Sergey Levine, a co-founder of Physical Intelligence, emphasizes that while current deployments remain experimental, the path to ubiquitous physical AI lies in creating systems that learn from continuous, real-world interaction, ultimately transforming how robots perceive and navigate the physical world.