ROBOTICS
Cooperative Multi-Robot Observation of Multiple Moving Targets
Simulation of multiple robots observing multiple moving targets. (Click on image for larger view) |
Challenging domains for multi-robot task selection are applications in which the proper actions by one robot are continuously dependent upon the current actions of other robot team members and the environment. To address this type of challenging domain, CESAR researchers have defined the CMOMMT application - the Cooperative Multi-robot Observation of Multiple Moving Targets. This application requires teams of robots with sensors of limited range to continually observe the motions of targets that are moving in a bounded area. The limited sensory range of the robots, the a priori unknown motions of the targets, and the possibility for many more targets than robots requires the robot team members to continually select their motions based upon the actions of their teammates and the target motions.
Optimal solutions to this problem are computationally intensive and cannot be implemented in real-time. CESAR researchers have therefore developed a distributed, heuristic, solution to the CMOMMT problem that performs well when compared to various control approaches. Our approach, which we call A-CMOMMT, involves the individual robot use of weighted local force vectors that attract them to nearby targets and repel them from nearby robots. The weights are computed in real time by a higher-level reasoning system in each robot, and are based on the relative locations of the nearby robots and targets. The weights are aimed at generating an improved collective behavior across robots when utilized by all robot team members.
We have implemented this approach in both simulation and on physical robots, and have performed extensive experimentation to determine the effectiveness of our hand-generated solution. We ran over 1,000,000 simulation test runs of up to 20 robots and targets, and over 800 physical robot experiments on our team of 4 Nomad robots. We studied both random and evasive target motions, and physical robot experiments in both cluttered and uncluttered environments. We compared the A-CMOMMT approach with a non-weighted local force vector approach, as well as two control cases in which robots either maintained fixed positions or are moved randomly.
Our results showed that the effectiveness of the A-CMOMMT solution is dependent upon a number of factors, including the relative numbers of robots and targets, the size of the work area, the motions of the targets (i.e., whether random or evasive), and the setting of the weights. In general, the A-CMOMMT algorithm performed best for a ratio of targets to robots greater than 1/2.
CESAR researchers have also identified the CMOMMT problem as an excellent domain for multi-robot learning. In continuing work (see section on multi-robot learning), we have developed approaches enabling robots to automatically learn efficient control approaches to solving the CMOMMT problem.
[