How far should I watch? Quantifying the effect of various observational capabilities on long-range situational awareness in multi-robot teams

Abstract

In our previous work, we showed that individual robots within a multi-robot team can gain long-distance situational awareness from passive observations of a single nearby neighbor without any explicit robot-to-robot communication. However, that prior work was developed only in simulation, and performance was not measured for real robot teams in physical space with realistic hardware limitations. Toward this end, we studied the performance of these methods in real robot scenarios with methods using more sophisticated techniques in machine learning to mitigate practical implementation problems. In this study, we further extend that work by characterizing the effects of changing history length and sensor range. Rather than finding that increasing history length and sensor range always yield better estimation performance, we find that the optimal history length and sensor range varies depending on the distance between the estimating robot and the robot being estimated. For estimation problems where the estimation target is nearby, longer histories actually degrade performance, and so sensor ranges could be increased instead. Conversely, for farther targets, history length is as valuable or more valuable than sensor range. Thus, just as optimal shutter speed varies with light availability and speed of the subject, passive situational awareness in multi-robot teams is best achieved with different strategies depending on proximity to locations of interest. All studies use the teams of Thymio II physical, two-wheeled robots in laboratory environments <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> . <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> Data and models used are available at https://github.com/PavlicLab/ACSOS2020_ReTLo_Extension.git.