Computer Science Department
School of Computer Science, Carnegie Mellon University
Handling Diverse Information Sources:
Paul E. Rybski, Manuela M. Veloso
To act intelligently in complex and dynamic environments, mobile robots must estimate the position of objects by using information obtained from a wide variety of sources, including different sensors, kinematic models, and communication from teammate robots. For any problem of reasonable complexity, mobile robots do not have the sensing capabilities necessary to simultaneously perceive all aspects of a dynamic environment, nor can they correct for possible unmodeled dynamics and noise. The "view" of the robot's sensors is hence narrow compared to the size of the environment. While the state of a single object is being updated, the evolving state of all other non-sensed objects must be predicted from models. In this paper, we formally describe the problem of estimating the state of objects in the environment where the robot can only task its sensors to view on object at a time. We contribute an object tracking method that generates, maintains, and selectively uses a disjoint space of hypotheses, whereby each hypothesis consists of a probabilistic state estimate that is generated by the individual sources of information. The priorities are set by the expected uncertainty in the object°«s motion/process model, as well as the uncertainties in the sources of information used to track their positions. Each individual robot guides its actions to track an object by selectively iterating through the multiple hypotheses sorted in terms of their expected state information. Our approach has been fully implemented in a team of AIBO soccer robots and used successfully in the RoboCup soccer competition. We describe the algorithm in detail and show extensive empirical results in simulation that demonstrate the effectiveness of our approach as well as an illustrative example from our actual robots.