Bayesian Mapping-Based Autonomous Exploration and Patrol of 3D Structured Indoor Environments with Multiple Flying Robots

Abstract

Mobile robots are frequently faced with mapping and exploring uncertain environments in surveillance, military, and convenience tasks. Often times, human teleoperation is either inconvenient or infeasible for these kinds missions. Furthermore, these tasks can be improved by cooperative multi-agent systems, where coordinating robotic efforts can be complicated and computationally-expensive. This paper presents a stochastic framework for autonomous exploration and patrol with multiple cooperating robots. The first contribution extends the authors’ prior work in single-robot exact occupancy grid mapping and autonomous exploration in a 2D environment to mapping and exploring in a 3D environment. The proposed 3D occupancy grid map is computed efficiently using an inverse sensor model that accounts for the sensor uncertainty, where we propose how several measurement sources may be fused together by considering depth readings individually. This approach is scalable to larger and more complex scenarios for real-time mapping. Furthermore, this paper shows how important aspects of a 3D map representing a structured environment are projected onto a 2D occupancy grid map, where an autonomous exploration algorithm is designed to select robotic motions that maximize map information gain. The mapping and exploration algorithms are demonstrated with an experiment where a quadrotor autonomously maps and explores an initially-uncertain environment. The second contribution is a novel approach to multi-vehicle cooperative patrol of environments based on map uncertainty. We propose a cooperative autonomous exploration algorithm, which applies a bidding-based framework to coordinate robotic efforts for improving occupancy grid map information gain. Since these exploration approaches are based on probabilistic knowledge about the map, the 3D occupancy grid map is systematically degraded over time to encourage the robots to revisit regions as time passes, thereby patrolling the environment. Furthermore, using a Bayesian framework and receding horizons, the algorithm is robust to dynamic obstacles within the mapping space. The efficacy of the proposed multi-vehicle cooperative patrol is illustrated with a simulation involving three robots patrolling a large floor plan with a non-cooperative person walking around the space.