I am interested in all aspects of the development of effective robot teams. To build effective teams we must address many challenges that are not a concern for single robot systems. Cooperation, diversity, communication, distributed learning and distributed planning are a few examples.
Distributed Planning and Communication
Many successful robot software systems are built in a layered fashion. Lower levels are concerned with sensor processing and motor actuation, while higher levels are focused on more abstract issues like behavior selection and/or planning. Our vision is to continue the layered approach above the level of the individual robot. Progressively higher levels will be concerned with the coordination of larger groups of agents. This distributed planning hierarchy is roughly equivalent to a military command structure: a theater commander might focus on tasking a division with high-level objectives (e.g. take and occupy city X) while field commanders face localized challenges (like defending and crossing bridges) in service of the higher-level goal. Similarly, planning and control of autonomous agents should be distributed so that planning is progressively more detailed and localized at lower levels. Real-time learning will be integrated at all levels to refine the planning process as agents gain experience in their environment.
Robot Team Diversity
We have developed a metric of robot team diversity that has been used experimentally in the evaluation of robots performing multirobot tasks (e.g. soccer, foraging and cooperative movement). It was discovered that behavioral diversity is an effective means for providing cooperation in multirobot teams but that the utility of diversity depends on the task. For instance, behavioral diversity seems to be useful in robot soccer teams, but not in foraging teams.
We now have a formal basis from which to investigate a number of important open issues relating to diversity. Why, for instance, is diversity important in some tasks, but not others? How does communication between agents impact the need for diversity? These are some of the issues we are investigating.
Social Potentials for Cooperative Behavior
The potential field approach is a well-known strategy for robot navigation. In this paradigm, repulsive and attractive fields are associated with important objects in the environment (e.g. goal locations or obstacles to avoid). To navigate, the robot computes the value of the vectors corresponding to each relevant field, then combines them (usually by summation) to compute a movement vector based on its current position. The result is emergent navigational behavior reflecting numerous c constraints and/or intentions encoded in the robot's task-solving behavior.
We have extended the mechanism to multiple robots so that the potential field impacting a robot's path is shaped by the presence of team or opponent robots. We call these potential functions "social potentials." This approach provides an elegant means for specifying team strategies in tasks like foraging, soccer and cooperative navigation.
This work is continuing as we seek to formalize and carefully analyze the various types of potentials appropriate for various multi-robot tasks.
|Research Interest Keywords|
|artificial intelligence, machine learning, mobile robots|
|The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.|
Contact Us | Update Instructions