A penalized batch-Bayesian approach to informative path planning for decentralized swarm robotic search

Swarm-robotic approaches to search and target localization, where target sources emit a spatially varying signal, promise unparalleled time efficiency and robustness. With most existing swarm search methods, it remains challenging to simultaneously preserve search efficiency and mathematical insight along with scalability and computational tractability. Our recently developed decentralized method, Bayes-Swarm-O, a model-based approach founded on batch Bayesian Optimization, has been shown to outperform state-of-the-art swarm heuristics in terms of search efficiency. However, this original Bayes-Swarm-O method did not account for the interactions between robots' decisions (aka samples in a batch) and was found to be sensitive to the prescribed balance between exploration and exploration. These limitations are alleviated in this paper, leading to significantly improved search efficiency and convergence, by respectively using a new marginalization penalization approach to embodied batch sampling and a dynamic adaptation of the exploration/exploitation balance during mission. In addition, this paper presents a systematic set of experiments executed through a new Pybullet-based distributed swarm search simulator, that analyzes the impact of increasing swarm size, partial peer observation, and choice of optimizer, on this updated algorithm, now called Bayes-Swarm-P. The advanced Bayes-Swarm-P method is also found to be clearly superior in terms of search efficiency and robustness when compared to three standard swarm search methods (namely Glowworm search, Levy walk, and exhaustive search) over simulated multimodal signal distributions and a skier/avalanche search and rescue problem.