Speaker: Drew Bagnell, Research Scientist, Robotics Institute
Date: Thursday, March 2, 2006
Abstract:
This presentation is based on joint work with Nathan Ratliff, Boris Sofman, Ellie Lin, Nicolas Vandapel, and Anthony Stentz
Programming behaviors for outdoor mobile robot navigation is hard. Machine learning promises to alleviate this difficulty but existing techniques often fall short. For instance, it is often the case that some features that, while potentially powerful for improving navigation, prove difficult to profit from as they generalize poorly to novel situations. Overhead imagery data, for instance, has the potential to greatly enhance autonomous robot navigation in complex outdoor environments. In practice, reliable and effective automated interpretation of imagery from diverse terrain, environmental conditions, and sensor varieties proves challenging. I'll discuss online, probabilistic models to effectively learn to use these scope-limited features by leveraging other features that, while perhaps otherwise more limited, generalize reliably.
I'll also discuss work on mobile robot learning based on demonstrated trajectories. This is a natural and potentially powerful approach to teaching a system. Unfortunately, most existing techniques to learn based on demonstrated trajectories face at least two important difficulties. First, it very hard to get "negative examples", in this framework; we can't actually drive the robot off a cliff or into a boulder. Secondly, it is very difficult to acquire long-horizon and goal-directed behavior by imitating a trainer. I'll talk about a new approach that addresses both concerns. It learns to map features of the world into costs for a planner in such a way so that resulting optimal plans mimic the trainer's behavior. This approach is powerful, as the behavior that a designer wishes the planner to execute is often clear, while specifying costs that engender this behavior is often much more difficult.
No comments:
Post a Comment