Integrated view and path planning for a fully autonomous mobile-manipulator system for 3D object modeling

We have designed and implemented a fully autonomous system for building a 3D model of an object in situ. Our system assumes no knowledge of object other than that it is within a bounding box whose location and size are known a priori, and furthermore, the environment is unknown. The system consists of a mobile manipulator, a powerbot mobile base with a six degrees of freedom (DOF) powercube arm mounted on it. The arm and the powerbot are equipped with line-scan range sensors, which provide range images that are used to build the work space Octree model and the object point cloud model. The object modeling system is comprised of three broad modules: (i) 3D model construction, (ii) a modeling view planner, which determines the next scanning pose for modeling, and (iii) a path planner, which determines a collision-free path to move the mobile-manipulator to a desired pose. Our research focuses on automating this process, which concerns the last two modules. Our modeling view planner calculates the target areas to be scanned, and then efficiently searches the five-dimensional viewpoint space to determine the best viewpoint for scanning the target areas. The path planning module itself consists of two subproblems: (a) exploration view planning, and (b) basic path planning. The former concerns developing an exploration strategy that facilitates manoeuvering the mobile-manipulator, for which we provide a formulation of C-space entropy reduction for range sensors for occupancy grid maps. For the basic path planning, a probabilistic Roadmap method (SBIC-PRM) is used for moving the mobile-manipulator to the view configuration. To construct a complete autonomous 3D modeling system, all three modules are repeatedly solved in an interleaved fashion. The process of modeling continues until a formally proved termination criteria is satisfied. We present extensive experimental results with our 3D object modeling system running on this real test-bed. The robot is started in unknown environments, and builds the object model in less than 35 scans. Each iteration takes about 4 minutes. The experimental results show the ability and efficacy of the system both in modeling the object and in the required exploration of the environment.