We have worked on a variety of planning problems for robots with very few sensors, very unreliable movements, or both. Results that use good algorithms to overcome such hardware limitations are useful both for reducing the cost and complexity of the robots we build, and for understanding the problems themselves with an eye toward solving them with more powerful robots.
Building on our prior work on robot localization with limited sensing, one recent series of results deals with navigation problems for nearly-sensorless mobile robots with bounded but substantial motion errors. The key observation is that certain kinds of environment features are useful for reducing uncertainty without explicit sensing. For example, we have shown that under the right conditions, a robot can localize itself in a corner of its environment, using only a noisy compass and a contact sensor. We have developed planning algorithms that exploit these kinds of strategies to enable long-range indoor navigation of very simple mobile robots.
One branch of this work deals with pursuit and evasion problems, in which a mobile robot pursuer moves to locate one or more unpredictably moving evaders. Our results show how to minimize the amount of time needed to locate the evaders in the worst case, and make significant improvements to the expected capture time when a probabilistic model for the evaders' movements is known.
We have also developed algorithms for target tracking by mobile robots with varying levels sensing ability, including scenarios in which the robot cooperates with an ambient sensor network in an energy-efficient way, and in which many robots work together to track a group of moving targets.
Much of our work is based on the idea of planning over a space of information states, which explicitly represent the robot's incomplete knowledge. However, for platforms in which there are strong limits on computation power (for example, due to limitations in space, weight, power, or cost), it may not be practical to compute those information states directly. As an alternative, we have proposed methods for maintaining low-complexity geometric representations of uncertainty that retain provable relationships (such as, for example, set containment) to the underlying "true" information state. Our results show that, in many cases, surprisingly simple representations are sufficient for the robot to make good decisions.
Because they can perform some mechanical tasks predictably and consistently, robots are well suited as part of an early intervention strategy for some autistic children, especially those who tend to perceive them as nonthreatening and intrinsically interesting. Research has demonstrated that robot-assisted autism therapy promotes increased speech and increased child-initiated interactions in children with Autism Spectrum Disorder (ASD). To make such robots more broadly accessible, we have designed and built a low-cost, interactive robot named CHARLIE (CHild-centered Adaptive Robot for Learning in an Interactive Environment). We have developed software for CHARLIE that autonomously detects and adapts to the perceived interest level of its user. To enhance the ability of robots like CHARLIE to respond to changes in their user's affective states, we are also developing new techniques for capturing breathing and heart rates remotely using a high precision, single-point infrared temperature sensor.Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.