Scaling up Data for Robot Learning through YouTube

A colloquium talk titled "Scaling up Data for Robot Learning through YouTube" presented by Stanford Assistant Professor Jeannette Bohg as part of the 2025 Winter Robotics Colloquium series at the Paul G. Allen School. Explore how robots can learn sensorimotor skills from human demonstrations, particularly through widely available videos of humans performing manipulation tasks. Professor Bohg discusses approaches toward enabling robots to learn by watching YouTube videos and outlines additional requirements for achieving generalist robots, including better policy architectures, multi-sensory data, online learning algorithms, and improved hardware. The talk addresses one of robotics' biggest challenges: equipping robots with universal sensorimotor skills needed for manipulating objects across diverse environments like homes, hospitals, warehouses, and factories. Professor Bohg, recipient of multiple Early Career and Best Paper awards including the 2019 IEEE Robotics and Automation Society Early Career Award, brings her expertise in perception and learning for autonomous robotic manipulation to this 71-minute presentation from March 7, 2025.