Leveraging low dimensional structure in motion planning and stochastic sequence analysis
Tuesday, April 17, 2012
Talk 4:30 pm
In both machine learning and motion planning, coping with the curse of dimensionality is of prime concern. However, although the theory and practice of dimensionality reduction in machine learning are quite advanced, the benefits of finding low dimensional structure in motion planning remain largely unexplored. In this talk, I will discuss my recent work on precisely identifying how low dimensional structure aids optimal motion planning, both through theory and experiments. I will show how these ideas lead naturally to the Learning Dimensional Descent algorithm, which achieves state-of-the-art performance on a real robot-arm-planning task.
The second part of this talk will show how these techniques are still applicable to problems that require reasoning about all paths, instead of just the optimal path. Such is the case in stochastic sequence analysis, where every path is assumed to occur with some probability, to be determined from empirical data. Experiments with human motion capture data will be shown demonstrating that this approach enables the efficient computation of probabilistic inferences that were previously intractable to compute. Connections to randomized and smooth motion planning will also be discussed.
Paul Vernaza is a postdoctoral fellow with the LAIRLab at Carnegie Mellon's Robotics Institute. He received the PhD from the University of Pennsylvania in 2011, working at the GRASP lab on problems at the intersection of machine learning and robotics with Prof. Dan Lee. Paul's research interests include exploiting tractable structure in high-dimensional motion planning and sequence analysis, developing physically inspired methods, and planning with topological constraints.