Learning 3-Dimensional Maps of Unstructured Environments on a Mobile Robot.

在移动机器人上学习非结构化环境的 3 维地图。

• 批准号: 5441387
• 负责人: Professor Dr. Andreas Birk
• 金额: --
• 依托单位: Chair of Computer Science
• 依托单位国家: 德国
• 项目类别: Research Grants
• 财政年份: 2005
• 资助国家: 德国
• 起止时间: 2004-12-31 至 2009-12-31
• 项目状态: 已结题
• 来源: https://gepris.dfg.de/gepris/projekt/5441387?language=en
• 关键词: Learning; Dimensional; Maps; Unstructured; Environments

The project deals with learning of a three dimensional representation of an unstructured indoor environment by an autonomous mobile robot. The general problem of mapping is a fundamental issue in robotics, as maps are crucial for the success of any non-trivial mission. In the past, significant efforts were devoted to build two dimensional maps like floorplans. These approaches simplify the problem, but they are well developed and they are sufficient for solving basic tasks. But mobile robots increasingly operate in three dimensions. This is due to significant advances in locomotion and the actual need to overcome for example slopes or stairs in any realistic environment. We propose an on-line learning algorithm, which creates a metric 3D representation encoded in the Virtual Reality Modeling Language (VRML) of an unstructured environment. The system is targeted for the real world, i.e., it has to cope with the unreliable and noisy navigation- and range-sensor data of a robot. The algorithm uses an evolutionary method where a VRML neighborhood model is extracted in realtime from a local 3D occupancy grid while the robot moves along. Inspired by previous successful work on realtime learning of 3D eye-hand coordination with a robot-arm, the evolutionary algorithm tries to generate VRML code that reproduce the vast amounts of sensor data. In doing so, we use a so-called neighborhood principle. Roughly speaking, the neighborhood principle uses a probabilistic estimation of the robots position in the world called the standpoint, which greatly reduces the complexity of learning steps. As demonstrator scenario, the system will be implemented on the IUB rescue robots.

该项目涉及由自主移动机器人学习非结构化室内环境的三维表示。绘制地图的一般问题是机器人技术中的一个基本问题，因为地图对于任何重要任务的成功都至关重要。在过去，人们花了大量的精力来绘制二维地图，比如平面图。这些方法简化了问题，但它们发展得很好，足以解决基本任务。但移动机器人越来越多地在三维空间中工作。这是由于运动方面的重大进步，以及在任何现实环境中克服斜坡或楼梯等实际需要。我们提出了一种在线学习算法，该算法以非结构化环境的虚拟现实建模语言（VRML）编码创建度量三维表示。该系统针对的是现实世界，也就是说，它必须处理机器人的导航和距离传感器数据的不可靠和噪声。该算法采用一种进化方法，在机器人移动时，从局部3D占用网格中实时提取VRML邻域模型。受先前成功的机器人手臂实时学习3D手眼协调工作的启发，进化算法试图生成VRML代码来复制大量的传感器数据。在这样做的过程中，我们使用了所谓的邻域原则。粗略地说，邻域原理使用了机器人在世界上的位置的概率估计，称为立场，这大大降低了学习步骤的复杂性。作为演示场景，该系统将在IUB救援机器人上实施。