Adaptive localization and mapping with application to planetary rovers

Future exploration rovers will be equipped with substantial onboard autonomy. SLAM is a fundamental part and has a close connection with robot perception, planning, and control. The community has made great progress in the past decade by enabling real-world solutions and is addressing important challenges in high-level scalability, resources awareness, and domain adaptation. A novel adaptive SLAM system is proposed to accomplish rover navigation and computational demands. It starts from a three-dimensional odometry dead reckoning solution and builds up to a full graph optimization that takes into account rover traction performance. A complete kinematics of the rover locomotion system improves the wheel odometry solution. In addition, an odometry error model is inferred using Gaussian processes (GPs) to predict nonsystematic errors induced by poor traction of the rover with the terrain. The nonparametric GP regression serves to adapt the localization and mapping to the current navigation demands (domain adaptation). The method brings scalability and adaptiveness to modern SLAM. Therefore, an adaptive strategy develops to adjust the image frame rate (active perception) and to influence the optimization backend by including high informative keyframes in the graph (adaptive information gain). The work is experimentally verified on a representative planetary rover under a realistic field test scenario. The results show a modern SLAM systems that adapt to the predicted error. The system maintains accuracy with less number of nodes taking the most benefit of both wheel and visual methods in a consistent graph-based smoothing approach.