Tracking an unpredictable target among occluding obstacles under localization uncertainties

• Published in: Robotics and autonomous systems 2002-01, Vol.38 (1), p.31-48
• Main Authors: Fabiani, P., González-Baños, H.-H., Latombe, J.-C., Lin, D.
• Format: Article
• Language: English
• Subjects: Game theory; Geometrical computing; Motion planning; Uncertainty; Visibility constraints
• ISSN: 0921-8890; 1872-793X
• DOI: 10.1016/S0921-8890(01)00170-1

This paper considers the problem of planning the motions of a mobile robot equipped with a visual sensor, whose task is to track an unpredictable moving object (called the target) in a workspace cluttered by obstacles. The planner must decide in real time how the robot should move in order to keep the target within its field of view. To do so, it must take into account the constraints imposed by obstacles to both visibility and motion. It must also deal with the uncertainties in both the robot’s position and the target’s future trajectory. This paper proposes a framework combining game theory and geometry to solve this multifold planning problem. At each time step, a probability distribution models the uncertainties associated to the robot and target localization. A utility function represents the reward associated with the possible goal states of the motion decision problem. This framework allows the simple modeling of specific tracking strategies, one of which was implemented and successfully tested with two mobile robots (one being the target). By simultaneously considering target visibility and position uncertainty, the robot takes advantage of landmarks scattered in the workspace to localize itself more precisely in order to track the target in the future better. Experiments have highlighted the relationship between the robot’s limited computing resources and the real-time constraints imposed by the tracking task. In particular, they reveal that there exists a planning horizon which achieves the best compromise between adaptivity and robustness in the robot’s behavior. This compromise strongly depends on the target’s escaping ability compared to the robot’s sensing, planning and moving capabilities. Symbolic inference tools could be used to adjust the planning horizon on-line.