Robotics Institute Seminar Series

Title: Where to Trust and How to Adapt Learned Models for Motion Planning

Abstract: The world outside our labs seldom conforms to the assumptions used to train our models. No matter how powerful our simulators or how big our datasets, our models will sometimes be wrong because they will eventually encounter states or environments that are outside the distribution of the training data. This talk will present our recent work on determining where learned models can be trusted and how to adapt learned models to new scenarios. Focusing on dynamics models and sampling distributions for trajectory optimization, these methods are designed for motion planning for a wide range of robotic systems. Our methods provide statistical guarantees on where learned dynamics models can be trusted and strong empirical performance on adapting sampling distributions for trajectory optimization to environments that are radically different from those used in training.