Evolving behaviour trees for supervisory control of robot swarms

Elliott Hogg,David Harvey,Sabine Hauert,Arthur Richards

doi:10.1007/s10015-020-00650-2

Abstract

Supervisory control of swarms is essential to their deployment in real-world scenarios to both monitor their operation and provide guidance. We explore mechanisms by which humans can provide supervisory control to swarms to improve their performance. Rather than have humans guess the correct form of supervisory control, we use artificial evolution to learn effective human-readable strategies. Behaviour trees are applied to represent human-readable decision strategies which are produced through evolution. These strategies can be thoroughly tested and can provide knowledge to be used in the future in a variety of scenarios. A simulated set of scenarios are investigated where a swarm of robots have to explore varying environments and reach sets of objectives. Effective supervisory control strategies are evolved to explore each environment using different local swarm behaviours. The evolved behaviour trees are examined in detail alongside swarm simulations to enable clear understanding of the supervisory strategies. We conclude by identifying the strengths in accelerated testing and the benefits of this approach for scenario exploration and training of human operators.

Highlights

Growing interest in the use of swarm systems has led to new questions regarding effective design for real-world environments [19]
This was achieved through observation of the swarm state and identifying the prior knowledge of the objective locations
We were able to explore a wide set of conditions that could be used to control the swarm and explore how this affects the types of strategies that are evolved

Summary

Introduction

Growing interest in the use of swarm systems has led to new questions regarding effective design for real-world environments [19]. Their use has great potential in areas ranging from search and rescue to automated agriculture, there are still few examples of real-world deployment. Human swarm interaction (HSI) aims to adapt swarm models into hybrid systems which operate with the aid of a human operator. The operator can compensate for the limitations of swarming behaviours and increase performance by interacting with the swarm.

Methods

Results

Conclusion