Affective Recruitment of Distributed Heterogeneous Agents

Aaron Gage, Robin R. MurphyAAAI 2004

Introduction

In the field of multi-robot systems, such as swarms of ground vehicles (UGVs) coordinated by unmanned aerial vehicles (UAVs), efficient task allocation is a critical challenge. Traditional recruitment methods, like greedy or random strategies, often struggle with scalability and communication overhead in decentralized environments. As the number of agents increases, the volume of coordination messages can saturate limited bandwidth, particularly in real-world scenarios like search-and-rescue where wireless connectivity is often unreliable.

This paper, presented by Aaron Gage and Robin R. Murphy, proposes a novel "affective recruitment" protocol. By integrating a formal model of the artificial emotion SHAME, the system allows robots to manage help requests autonomously, significantly reducing communication traffic while maintaining robustness to message loss.

The Affective Recruitment Protocol

The core of the proposed system is an extension of the ALLIANCE architecture, utilizing a SHAME variable ($s$) to regulate agent responses. Unlike greedy algorithms where every idle robot immediately bids on a task, the affective protocol introduces a threshold-based delay.

The recruitment protocol utilizes a multi-step handshake to ensure reliable coordination between requester and responder robots.

The interaction follows a specific logic:

  1. SHAME Accumulation: When a requester broadcasts a HELP message, potential responders do not necessarily answer immediately. If a robot's internal SHAME level is below a threshold ($t$), it ignores the request. However, ignoring a request causes the SHAME level to increase, with higher increments for robots closer to the requester.
  2. Threshold and Acceptance: Once SHAME exceeds the threshold, the robot sends an ACCEPT message.
  3. Handshake: A three-way handshake (Help-Accept-Arrival/Agree) ensures that recruitment succeeds even in noisy environments.
  4. Decay and Reset: SHAME levels decay over time if no requests are active and reset to zero upon the successful completion of a task.

Experimental Evaluation

The researchers evaluated the affective strategy through 66 simulated trials and a physical proof-of-concept using ATRV-Jr robots. The simulation environment consisted of a 100x100 unit grid representing a minefield, with a team comprising one UAV requester and up to 22 UGV responders.

The simulation environment tracks the "Shame over Time" for individual robots, demonstrating the emotional buildup required before a recruitment commitment.

Results and Discussion

The affective strategy demonstrated significant advantages in communication efficiency and scalability:

  • Message Economy: The protocol required 19.1% fewer messages overall compared to greedy and random methods. This advantage became more pronounced as team sizes grew; with 20 idle robots, the affective strategy sent 36.3% fewer messages than the greedy approach.
  • Robustness: The system remained efficient even under a 25% message loss rate. The repetitive nature of the HELP broadcast ensures that as SHAME rises, a robot will eventually hear and respond to a request, making the system naturally resilient to dropped packets.
  • The Patience Trade-off: The primary cost of this efficiency is increased wait time. Greedy recruitment was approximately 39.4% faster because it lacks the "emotional build-up" period. However, in bandwidth-constrained environments, this trade-off is often necessary to prevent network collapse.

Conclusion

Affective recruitment provides a scalable, low-bandwidth solution for coordinating heterogeneous robot teams. By treating communication as a filtered process rather than a constant broadcast, the system mimics social pressures to optimize network usage. Future work will focus on automated parameter tuning via machine learning and integrating more complex emotional models, such as "guilt," to handle task preemption and better match robot capabilities to specific mission requirements.

Summary updated

BibTeX
@inproceedings{gage2004affective,
  author = {Aaron Gage and Robin R. Murphy},
  title = {{Affective Recruitment of Distributed Heterogeneous Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2004},
}
Figure