Observational Uncertainty in Plan Recognition Among Interacting Robots

Plan recognition is the process of observing another agent's behavior(s) and inferring what, and possibly why, the agent is acting as it is. Plan recognition becomes a very important means of acquiring such information about other agents in situations and domains where explicit communication is either very costly, dangerous, or impossible. Performing plan recognition in a physical domain (i.e. the real world) forces the world's ubiquitous uncertainty upon the observing agent because of the necessity to use real sensors to make the observations. We have developed a multiple resolution, hierarchical plan recognition system to coordinate the motion of two interacting mobile robots. Uncertainty arises in the system from dead reckoning errors that accumulate while the robots are moving, as well as by errors in the computer vision system that is used to detect the other agent's behaviors. Based upon belief networks, the plan recognition system gracefully degrades in performance as the level of uncertainty about observations increase.