Unlocking Scalable Robot Learning in the Real World

This Allen School Colloquia Series talk features Karl Pertsch from UC Berkeley and Stanford University discussing the challenges and solutions for scaling robot learning in real-world environments. Explore how robotics differs from other machine learning domains and learn about approaches that have enabled the creation of the largest robot learning datasets to date. Discover how generalist robot policies can be developed to perform various manipulation tasks in unfamiliar environments through natural language prompting. The presentation covers three key elements of modern machine learning pipelines: data, models, and evaluations, while addressing current limitations and open challenges in developing general robot control policies. Karl Pertsch, a postdoc at UC Berkeley and Stanford working with Sergey Levine and Chelsea Finn, shares insights from his award-winning research, including work that received the Best Conference Paper Award at ICRA'24 and two Outstanding Paper Award Finalists at CoRL'24.