复杂动态环境下基于滑移参数的机器人地面分类问题研究

The accurate and real-time terrain classification is a prerequisite for mobile robots to avoid non-geometric hazards, and also an important means to improve the performance of robot pose estimation and motion control. In this project, the problem of robotic terrain classification based on slip parameters is investigated for wheeled/tracked mobile robots working in the complex and dynamic field environment. Specifically, considering the non-line-of-sight effect and magnetic field distortion in the complex environment, the adaptive joint estimation method of robot pose and slip parameters will be studied. Considering the potential class evolution and class conditional probability drift in the dynamic environment, the hierarchical ensemble learning method based on binary classifiers will be studied. Considering the lack of labelled terrain perception data, this project will study how to improve the performance of terrain classification based on the manifold hypothesis and multi-view disagreement. This project concentrates on improving the perception ability of mobile robots to the unknown terrain environment without human instructions, which could inspire the research of autonomous robots.

准确实时的地面分类是移动机器人避免非几何危害的前提，也是提升机器人位姿估计、运动控制等方面性能的重要手段。本课题针对在复杂动态的野外环境下工作的轮式/履带式移动机器人，研究基于滑移参数的机器人地面分类问题。具体地，针对野外复杂环境普遍存在非视距效应与磁场畸变的现象，研究机器人位姿与滑移参数的自适应联合估计方法；针对动态环境中潜在的类演化与类条件概率漂移问题，研究基于二元分类器的分层集成学习方法；针对有标注环境感知数据缺乏的问题，研究基于流形假设与视图分歧的地面分类性能提升方法。本课题旨在提升移动机器人在无人指导条件下对未知开放地面环境的感知能力，为自主机器人的研究提供了新的思路。

准确实时的地面分类是移动机器人避免非几何危害的前提，也是提升机器人位姿估计、运动控制等方面性能的重要手段。本项目针对在复杂动态的野外环境下工作的轮式/履带式移动机器人，从机器人滑移参数在线估计、地面分类模型初建和更新三个方面，系统性地研究了基于滑移参数的机器人地面分类问题。项目执行期间，在IEEE汇刊等SCI索引期刊上发表学术论文11篇；授权发明专利7项，授权软件著作权2项；培养相关领域在读硕士研究生2名；获得第八届安徽省专利奖银奖、第二十三届中国专利奖优秀奖；获得中国科学技术大学墨子杰出青年特资津贴一等资助；获得国家自然科学基金面上项目资助。