 |
| Typical small UAV (AscTec Pelican) |
Coordinating a team of agents, which could be a search team, cleaning robots, flying drones for surveillance or environmental monitoring is a highly relevant problem. If the environment is unknown or subject to change, an a priori planning algorithm becomes difficult to apply. Therefore we looked into decentralized self-organizing algorithms to do the job.
In a joint work with István Fehérvári, Evsen Yanmaz and Wilfried Elmenreich (me), we evolve controllers for a team of unmanned aerial vehicles (UAVs) with the task to observe or cover a partially obstructed area.
The respective agents are limited in their sensory inputs to local observations of the environment without the ability to determine their absolute position or those of others. Each agent is equipped with a number of sensors that can detect the presence of other agents, an obstacle and the border of the area.
 |
| Simulation and evaluation model |
The controller of an agent is implemented as an artificial neural network. The fitness for a given configuration is derived from the average spatial coverage over several simulation runs. The area coverage performance of the evolved controllers with different number of sensors is compared to reference movement models like random walk, random direction, and an algorithm based on the belief of the intention of agents met during the execution of the simulation. Our results show that evolved controllers can create a self-organizing cooperating team of agents that exploit the advantages provided by their sensors and outperform naïve coverage algorithms and also reach the performance of a recent algorithm that is using additional information as well.
The work was presented in a talk at the European Meeting on Cybernetics and Systems Research (EMCSR 2012) in Vienna, Austria. Slides are available via slideshare:
No comments:
Post a Comment