
Robot Skill Acquisition: Policy Representation and Data Generation
Shuran Song of Stanford University argues that the bottleneck in robot learning is not algorithms or hardware but data itself, and specifically data that is scalable, reusable, and what she calls robot-complete. Delivered as part of Stanford's AA289 Robotics and Autonomous Systems Seminar, the talk examines why simply scaling up dataset size, the strategy that worked for other machine learning fields, has not translated cleanly into robotics. Song walks through competing approaches to generating training data for robot policies, pointing out hidden costs that each one carries, from collection effort to poor transfer across tasks and hardware. She then outlines directions for sidestepping these obstacles so that robot learning can benefit from the same scaling trends seen elsewhere in machine learning. The talk centers on policy representation and data generation as the two levers most likely to move the field forward, grounded in her own research on robot learning systems.