服务机器人广域环境感知和场景认知理论及应用

The effective enhancement of perception and cognition capability in robotics is a reliable safeguard of high-quality service of the robots in wide-area environments. This proposal aims at extensive research on basic theories and key technologies of wide-area environmental perception and scene cognition for robots, on the basis of integrative utilization of multiple theories and disciplines such as robotics, artificial intelligence, machine vision, machine learning, and cognitive science. The main research contents are listed as follows. Firstly, we will design a convolutional neural network with a dynamic property, and implement wide-area feature binding and integration by self-organized neuronal synchronization. Secondly, by combining the addressable memory and deep neural network, we will establish a differentiable computing architecture to imitate the human ability of associative memory, thereby achieving the scene space cognition based on information recovery. Thirdly, we will build a non-analytical pose measurement model based on the deep learning framework, and seek to break the bottleneck in which the traditional visual measurement techniques hardly deal with the irregular objects. Fourthly, we will integrate the topology modeling and local rigidity field to describe the dynamic wide environments. At last, the effectiveness and applicability of the proposed theories and methods are verified by indoor and outdoor service applications and demonstrations on the self-developed service robot platform. The scientific significance of this proposal is to explore bio-inspired perception and cognition mechanisms through interdisciplinary efforts, thus offering important theoretical basis and technical support for the enhancement of perception ability and autonomy of service robots.

机器人感知与认知能力的有效提升是广域环境下机器人高质量服务作业的可靠保障。本项目综合机器人学、人工智能、机器视觉、机器学习、认知科学等多学科的理论和方法，针对机器人广域环境感知和场景认知的基础理论和关键技术开展深入研究。主要研究内容包括：设计具有动力学特性的卷积神经网，以神经元自组织同步发放的方式实现大范围特征绑定与整合；结合可寻址存储器和深度神经网，建立可微分的计算架构模拟人类联想记忆，实现基于信息修复的场景空间认知能力；基于深度学习架构建立非解析式位姿测量模型，突破传统视觉测量技术无法应对不规则物体位姿测量的瓶颈；通过整体拓扑与局部硬度场的混合，实现广域动态环境建模；在此基础上，通过室内、室外动态环境下机器人的服务示范，对所提理论和方法的有效性和适用性进行验证。本项目的科学意义在于通过多学科交叉融合来探索仿生感知和认知机理，为提升服务机器人的感知和自主能力提供重要的理论参考和技术支撑。

本项目针对机器人广域环境感知和场景认知的基础理论和关键技术开展深入研究。首先，提出了一种基于弱监督策略的小样本注意力学习方法，在获得对潜在目标选择性注意的同时，实现对指定目标的高效编码。其次，通过对多源信息的绑定，实现了场景结构特征提取，并以曲线的形式输出，机器人进而实现通道与障碍的感知；还面向动态场景，建立了基于随机采样的概率场模型，开发了动态物体挖掘算法。第三，模仿人类导航，对“路径”概念进行了分析，建立了路径感知量的可计算模型。在此基础上，设计了一种隐式记忆路径的深度编码器，将感知结果映射为态势值，为机器人推理式导航提供依据。第四，给出了一种基于多任务深度卷积神经网络的抓取位置检测方法，并将轻量级网络和注意力机制融入全局卷积语义分割网络中实现抓取位置检测。最后，提出了机器人广域环境拓扑建模的方法，通过人工传授与自主探索相结合的方式构建环境拓扑地图进而实现大范围导航。搭建了主动视觉平台、分布式传感网、操作型移动机器人、多机器人协同系统等实验平台，并进行了实验验证。发表期刊论文48篇，会议论文18篇，其中IEEE Transactions 论文28篇。申请发明专利6项、授权发明专利10项，软件著作权登记6项。培养青年科技骨干6名，培养博士研究生3名、硕士研究生4名。