We look forward to seeing you online on Thursday, 3/30 for the Vision Speaker Series.
Imitation is an effective and safe paradigm for teaching machines new skills. In this talk, I will explore novel frameworks for imitation learning of robust sensorimotor policies for autonomous and assistive systems. First, I will present a large-scale approach for imitation learning from observations for autonomous vehicles. I will demonstrate how the proposed approach can facilitate learning of generalized models that are platform and settings-agnostic, and can robustly operate across dynamic urban settings. Nonetheless, I will show that state-of-the-art models still struggle when learning basic navigation skills in such complex settings. To address this difficult sensorimotor learning task, I will introduce a novel knowledge distillation framework for scaffolding and coaching the training of sensorimotor agents. The effective vision-based behavior cloning approach improves upon prior results by over 20.6% in driving score in CARLA, without requiring LiDAR, historical observations, an ensemble of models, or on-policy learning. These two results demonstrate an (urgent) need to fully explore the limits and limitations of imitation learning, for which we have only begun to scratch the surface.
Eshed Ohn-Bar is an Assistant Professor with appointments in the ECE and CS departments at Boston University. Eshed’s research lies at the intersection of machine intelligence, computer vision, and systems for human-machine interaction (particularly for accessibility). Prior to BU, he was a Humbodlt Fellow at Max Planck Institute for Intelligent Systems. His work has received the 2017 best PhD dissertation award from the IEEE Intelligent Transportation Systems Society, the best paper award at the the 2019 Web for All conference, honorable mention for the best student paper award at the 2014 and 2016 International Conference on Pattern Recognition, and best paper award at the Workshop on Analysis and Modeling of Faces and Gestures at CVPR 2013. His team was also a semi-finalist at the 2022 Department of Transportation’s Inclusive Design Grand Challenge. Eshed received the BS degree in Mathematics from UC Los Angeles in 2010, MEd from UC Los Angeles in 2011, and the PhD degree in Electrical Engineering from UC San Diego in 2017.