Adaptive Selection of Informative Path Planning Strategies via Reinforcement Learning

In our previous work, we designed a systematic policy to prioritize sampling\nlocations to lead significant accuracy improvement in spatial interpolation by\nusing the prediction uncertainty of Gaussian Process Regression (GPR) as\n"attraction force" to deployed robots in path planning. Although the\nintegration with Traveling Salesman Problem (TSP) solvers was also shown to\nproduce relatively short travel distance, we here hypothesise several factors\nthat could decrease the overall prediction precision as well because\nsub-optimal locations may eventually be included in their paths. To address\nthis issue, in this paper, we first explore "local planning" approaches\nadopting various spatial ranges within which next sampling locations are\nprioritized to investigate their effects on the prediction performance as well\nas incurred travel distance. Also, Reinforcement Learning (RL)-based high-level\ncontrollers are trained to adaptively produce blended plans from a particular\nset of local planners to inherit unique strengths from that selection depending\non latest prediction states. Our experiments on use cases of temperature\nmonitoring robots demonstrate that the dynamic mixtures of planners can not\nonly generate sophisticated, informative plans that a single planner could not\ncreate alone but also ensure significantly reduced travel distances at no cost\nof prediction reliability without any assist of additional modules for shortest\npath calculation.\n