Skip to content. Skip to navigation
CIM Menus
 

School of Computer Science

Active Semantic Robotics


Dave Meger, Tomlinson Postdoctoral Scholar
Mobile Robotics Lab McGill School of Computer Science

April 9, 2015 at  10:00 AM
McConnell Engineering Room 437

Abstract:

My research objective is to enable robots to utilize increasingly rich, robust, and adaptable representations of their environment. I believe this crucially requires the combination of perception, planning and control, and have focused on algorithms that jointly reason about 2 or more of these factors.

First, I will describe my work on 3D Scene Understanding, where the world is perceived as a collections of objects, along with their geometry and inter-relations. This model allows for strong spatial priors to be applied during perception and fuses information from multiple sensors and viewpoints. I have developed an efficient sampling-based inference approach for this problem, which allows for its effective use on a robot system. Visual occlusion is a key concern for any image-based model, and represents a point of strength for the 3D scene representation. I have shown how to learn occlusion-aware parts-based object models and how to apply these during perceptual inference to achieve strong improvements on recognizing occluded object instances. Second, I will describe my work on planning under uncertainty, both in the context of object recognition models and also for geometric mapping. Finally, I will describe the recent use of reinforcement learning methods to control the Aqua robot while swimming, enabling acrobatic motions to be completed with very little human engineering effort.

Bio:

David Meger is currently a Tomlinson Postdoctoral Scholar at McGill's Mobile Robotics lab in the School of Computer Science. He received a BSc from UBC in 2004 and MSc from McGill in 2006. His PhD thesis, supervised by Dr. Jim Little at UBC, was titled "Visual Object Recognition for Mobile Platforms." David served as co-chair for the Computer Robot Vision conference in 2013 and 2014. His research involves algorithms for perception and control that have been demonstrated on many successful robotic systems including Curious George, the three-time winner of the Semantic Robot Vision Challenge, an international robotic object recognition contest.