Who doesn't love a beautiful, accurate, and detailed 3D model built by a robot? And who doesn't love the challenge of planning and following trajectories through high dimensional state spaces, a challenge enabled by these maps?
The result of these approaches--- far too often-- is a slow, brittle, error-prone robot. For example: a data association error can cause the map to fail catastrophically; planning in high-dimensional state spaces is very slow (best illustrated, perhaps, by the now-ubiquitous "10x" speedup used in so many demonstration videos!); incorrectly estimating the shape of a translucent bottle leads to a failed grasp.
In many common cases, contemporary systems can't compete with simple heuristics; for example, the easiest and most reliable way to get to the kitchen might be to follow the wall on the left until the robot reaches the second door; a grasp can often be achieved reliably by visual servoing and closing a gripper until a specified resistance is encountered.
In this talk, I'll describe how we're trying to build reliable, simple, and fast robots. We don't abandon our modern methods--- in difficult cases, a high-dimensional planner operating on a 3D detailed model is the best way. Instead, we see the problem as one of introspection: can the robot determine *when* it should use a simple method versus a more complex one? We'll elaborate on these ideas and our initial efforts and results.
Edwin Olson is an Associate Professor of Computer Science and Engineering at the University of Michigan. He is the director of the APRIL robotics lab, which studies Autonomy, Perception, Robotics, Interfaces, and Learning. His active research projects include applications to explosive ordinance disposal, search and rescue, multi-robot communication, railway safety, and automobile autonomy and safety.
In 2010, he led the winning team in the MAGIC 2010 competition by developing a team of 14 robots that semi-autonomously explored and mapped a large-scale urban environment. For winning, the U.S. Department of Defense awarded him $750,000. He was named one of Popular Science's "Brilliant Ten" in September, 2012. In 2013, he was awarded a DARPA Young Faculty Award.
He received a PhD from the Massachusetts Institute of Technology in 2008 for his work in robust robot mapping. During his time as a PhD student, he was a core member of their DARPA Urban Challenge Teamwhich finished the race in 4th place. His work on autonomous cars continues in cooperation with Ford Motor Company on the Next Generation Vehicle project.
He is active in the open source software community as one of the original developers of the message-passing system LCM, and the creator of the OrcBoard robotics controller. Much of his current software is available under open source licenses.Headshot: http://april.eecs.umich.edu/people/ebolson/headshots/DSC_6746.JPG