
Robot Learning in the Era of Large Pretrained Models
Dorsa Sadigh of Stanford University presents this AA289 Robotics and Autonomous Systems Seminar talk on how interactive robot learning can draw on large pretrained models. She covers two main threads: the role of pretraining in learning visual representations, including how language can ground those representations for downstream robotics tasks, and the role of dataset choice during pretraining, including how to guide large scale data collection and what counts as high quality data for imitation learning. She discusses recent work on guiding data collection to improve compositional generalization in learned policies, and closes with examples of tapping the context stored in large language models and vision-language models for robotics applications. The talk is aimed at a technical audience familiar with machine learning and robotics fundamentals, and is part of Stanford's AA289 seminar series.