Value Iteration Networks with Double Estimator for Planetary Rover Path Planning

Path planning technology is significant for planetary rovers that perform exploration missions in unfamiliar environments. In this work, we propose a novel global path planning algorithm, based on the value iteration network (VIN), which is embedded within a differentiable planning module, built on the value iteration (VI) algorithm, and has emerged as an effective method to learn to plan. Despite the capability of learning environment dynamics and performing long-range reasoning, the VIN suffers from several limitations, including sensitivity to initialization and poor performance in large-scale domains. We introduce the double value iteration network (dVIN), which decouples action selection and value estimation in the VI module, using the weighted double estimator method to approximate the maximum expected value, instead of maximizing over the estimated action value. We have devised a simple, yet effective, two-stage training strategy for VI-based models to address the problem of high computational cost and poor performance in large-size domains. We evaluate the dVIN on planning problems in grid-world domains and realistic datasets, generated from terrain images of a moon landscape. We show that our dVIN empirically outperforms the baseline methods and generalize better to large-scale environments.