This was a hybrid event with in-person attendance in Levine 307 and virtual attendance…
Decades of rigorous research in dynamical systems and control helped us integrate robots into a wide variety of domains, ranging from factory floors to the moon. Today, it would appear that deep learning has taken over the torch and will bring robots to our homes, freeing us all from banal chores. In this utopian vision, learning-based approaches tend to replace analytical methods. Moving away from handcrafted bespoke solutions to generalist robots that can operate in unstructured environments. But one can instead view learning-based and analytical approaches as two ends of a broad spectrum, with one end optimizing for reliability (at the cost of human effort) and the other for emergent intelligence (at the cost of data and computation). In this talk, I will argue why it is better for robots to be in the middle of this broad spectrum. Using manipulation as a case study, I will discuss how our lab combines ideas from dynamical systems and machine learning to overcome three often-overlooked issues with contemporary methods: i) high barrier to entry due to demands for expensive computational resources and annotated data, ii) inability to handle new tasks without relying on significant user expertise (e.g., for reward or controller design, hyperparameter tuning, data collection and curation), and iii) unreliable behaviors due to inscrutable and unpredictable learned policies. Addressing these issues will enable robot learning to escape the confines of well-resourced research labs and positively impact the larger society.