Can One AI Model Control Every Robot? | Sergey Levine

Physical intelligence—the ability of AI to control robots and navigate the physical world—remains a significant hurdle despite advancements in cognitive AI. While language models excel at symbolic tasks like programming or math, they lack the embodied experience required for everyday physical actions like folding laundry or manipulating objects. Sergey Levine, co-founder of Physical Intelligence, posits that a general-purpose robotic foundation model, rather than specialized, domain-specific systems, will ultimately solve this. By training models on diverse physical interactions, these systems develop a fundamental understanding of cause and effect, enabling them to generalize across various hardware platforms. This transition from rigid, programmed automation to learning-enabled, adaptive robotics mirrors the evolution of general-purpose software, potentially leading to ubiquitous, intelligent physical agents that operate across homes and factories with human-like common sense and adaptability.