Learning-based Real-time Down-sampling for Scalable Decentralized Decision-Making in Bayes-Swarm Search

A swarm of robots working together promise efficient solutions towards searching for signal sources in hazard localization and search-&-rescue applications. An effective class of such swarm search algorithms is one that constructs, progressively refines and uses a belief model of the signal environment. These algorithms however risk becoming too computationally burdensome for real-time or onboard decentralized implementation as the size of the swarm or mission duration increases. Other applications that involve collecting increasing amount of spatially-distributed data and building belief models thereof to make decisions face this risk as well. To alleviate this risk, this paper presents a new learning based down-sampling approach applied to a notable search algorithm called Bayes-Swarm, which uses a Gaussian Process model of the signal environment. The proposed down-sampling approach involves three key elements: 1) new abstraction of the input encoding visited locations, and the down-sampled data; 2) a probabilistic loss function based on Sinkhorn distance to capture the down-sampling performance; and 3) a convolutional neural network (CNN) to perform this down-sampling. These elements enable learning a generalized way to down-sample observations without requiring expensive simulations of the swarm search process. The new Bayes-Swarm implementation with the CNN-based down-sampling (CNN-Bayes-Swarm) is found to be agnostic to increasing data set size unlike original Bayes-Swarm, without significant compromise in search completion time. CNN-Bayes-Swarm also outperforms other baseline down-samplers, in terms of search performance and computing costs considered together.