Multi-agent recruitment * distributed team of heterogeneous robots (UGVs, UAV) * robots must cooperate (UAV finds interesting target, UGV requested to investigate)

Domain constraints * Wireless (shared, finite, unreliable) * Team issues: decentralized, distributed. Variable size. * No task pre-emption * No knowledge of future tasks * Hybrid behavior-based architecture. No modeling of other agents, no shared representation/model

Approach * Contract Net Protocol with first-price auction, plus emotional motivation * Robots only respond to announcements given sufficient motivation. Saves communication overhead, scales better, reduces power consumption, stealth. * Persistent motivation leads to uniform recruitment (fairness). Robot that doesn't help now will help later.

Shame * each robot has shame based on degree to which robot has not contributed to needs of team. * standards-based emotion from OCC model * shame increases when they ignore bid

Emotional model, parameters * Shame s, where 0<=s<=1 * Threshold t where t<=1 * c: constant added to s when HELP ignored * d(D): distance

Experiments * compare affective recruitment (1/D, 1/D^2) against greedy and random strategies (team size, random message loss, shortcomings of greedy, fairness). * metrics: time elapsed between request and robot arrival, messages transmitted * domain: distributed mine detection task (NAVSEA Coastal Systems station). UAV and UGVs. UAV performs raster scan over minefield, requests UGV to investigate. * Java/JINI * Variable size of robot team (4-53 UGVs)

Experimental results (team size) * Greedy scheduler scales linearly in messages transmitted with team size * Affective scales at less than linear * Affective recruitment takes longer for recruitment to start (shame has to build up) * Greedy faster than affective, but as team gets large enough the two will converge.

Experimental results (fairness) * affective more fair, greedy tends to favor a robot

Summary * use communications channel more efficiently (35% savings in coms) * affective approach requires more time than greedy. Degree of difference depends on parameters as greedy is a special case of affective. * tested on real UGVs. * first known application of emotional model to multi-robot task allocation. * Communication overhead scales better than O(n). * less sensitive to task order * tolerant of communication failures * fairness