Statistical Methods for Collaborative Multi-Robot Mapping

点击翻译按钮获取中文摘要

Abstract IIS-9877033Thrun, SebastianCarnegie Mellon University $69,602 - 12 mos.STATISTICAL METHODS FOR COOPERATIVE MULTI-ROBOT MAPPINGThis is the first year funding of a three year continuing award. This project will devise a new family of statistical algorithms that enable teams of mobile robots to jointly acquire maps of indoor environments. The majority of successful mobile robot applications todate rely on maps; yet, the problem of automatically acquiring such maps with one or more robots remains largely unsolved. To build such maps, three sources of uncertainty must be overcome: uncertainty in sensing (e.g., with range finders), uncertainty in robot odometry, and uncertainty in the initial poses of the different robots relative to each other. In addition, robots must integrate information gathered distributedly, and they must coordinate their exploration strategies.The research is based on a unifying statistical framework, recently devised by the PI and others. This framework poses the problem of building maps as a maximum likelihood problem, of finding the most likely map given the data acquired by a team of robots. It uses efficient statistical techniques (specifically Dempster's EM algorithm) for likelihood maximization. The project goes beyond previous work in several aspects: First, it will address the problem of team-based map acquisition, where robots have to determine their relative location to each other in order to build a single map. This will be achieved by introducing additional degrees of freedom into the basic statistical framework. Second, maps will be represented by conjunctions of atomic object shapes (eg, rectangles, circles or segments thereof), instead of grids or similar non-iconic representations. To estimate the total number, sizes and poses of suchobjects, the project will apply fast, sampling-based algorithms. Third, the project seeks to devise algorithms that enable teams of robots to collaboratively explore an environment. This will be achieved by methods for maximizing information gain, as measured by expected relative entropy. Finally, this project will pursue explorative research on learning maps in dynamic environments. If successful, this project is expected to lead to a new, more robust and scalable family of algorithms for mobile robot mapping.