A New Path to Embodied AI

Building general-purpose embodied intelligence has been a core goal of AI since its inception 70 years ago. The field has evolved through multiple hypotheses, beginning with search as a solution, followed by knowledge-based systems, and, most recently, leveraging large language models for robotics. Yet, the challenge of creating generalist embodied agents capable of performing thousands of tasks across diverse environments remains unsolved. Building such a generalist robotic agent presents a "chicken-and-egg" problem: to train agents for generalization, we need vast amounts of robotic/agent data from varied environments, but gathering such data is impractical without deploying robots/agents that already generalize. I propose a new framework that leverages alternative sources of "indirect" supervision to build robotic agents from the ground up. I'll focus on three key principles: (1) How to generate self-supervised data for training robots? (2) How to bootstrap robot learning by observing human videos? and (3) How can robots learn to adapt via large-scale training in simulation? I'll demonstrate this framework's potential for scaling robotics AI through several case studies.