基于深度学习的起重机器人优良位形映射机制研究

How to obtain good picking/placing configurations is a key issue of lift path planning whose initial and goal are specified by poses of lifted objects. Most existing approaches solved this problem by treating it as an IK problem. These approaches did not guarantee that the obtained configuration is good for generating a high quality path because they just took kinematical constraints into account. Instead, inspired by the motion of human picking one object in clutter environments, this proposal treat the issue as a mapping from lifting situation to good configurations space, and our previous works showed that the mapping could be learned from examples. So, this proposal will attempt to study the mapping using deep learning. We firstly study on the representation of lifting situations and use unsupervised learning to fetch the high level features of 3D lifting environments. And then we will investigate the underlying statistics of the two modalities—3D lifting environment and lifting task description, and propose a feature learning model for the two modalities. On this basis, the proposal will develop a deep network to realize a predictor for good configurations and its training methods. After training, once a predictor is obtained we can analyze its interior structure by visualizing its network parameters and its hidden layers. This analysis will help us have a more profound understanding of the mapping from lifting situations to good configurations. It will also provide the theoretical basis for optimizing the deep network or redesigning a more accurate predictor. The expected research achievements of this proposal will put forward the development of autonomy of a crane robot. In the meantime, the research achievements also will provide a new perspective for manipulation planning of redundant mobile manipulators.

优良位形获取是被吊物位姿给定的吊装路径规划中一个核心问题，现有研究通常单纯地将其看作逆向运动学问题，因其仅考虑运动学约束等局部因素而无法确保所得位形优良。受人类日常抓取操作启发，本项目从更宏观的视角将此问题看作吊装情形到优良位形的映射，并通过前期预研发现可望通过历史数据学习到此映射进而获得优良位形，为此尝试利用深度学习技术研究吊装情形到起重机器人优良位形的映射机制。首先探讨吊装情形的表示，研究基于深度无监督学习的三维环境特征提取；然后分析吊装任务和三维环境的内在特性，提出相应多模态数据处理方法；接着设计优良位形预测器的深度网络结构及其训练策略，以获得表达吊装情形到优良位形映射机制的预测器；最后，分析所得预测器的内部结构，可望对已获映射机制有更深刻而系统的理解，为优化深度网络或重新设计更准、更强的优良位形预测器提供理论依据。本研究将推动起重机器人自主吊装的发展，也将为抓取规划研究提供新视角。

运动规划代表着起重机器人智能水平，而优良位形获取是其中一个核心问题，现有研究通常将其看作逆向运动学问题来解决，因其仅考虑运动学约束等局部因素而无法确保所得位形优良。受人们日常抓取操作启发，本项目从另一个新视角将此问题抽象为吊装情形到优良位形的映射问题，并研究利用深度学习技术获得此映射，获得如下研究成果：1）探明了被吊物起始位姿、终止位姿主要通过两位姿相对信息影响优良位形，并给出了基于极坐标的吊装任务信息的表达方法；提出了基于二维栅格的吊装障碍物环境的表示方法；2）提出了基于起重机器人站位的优良位形区的表示方法，使得优良位形预测器更容易训练。3）获得了一个优良位形区预测器，并据此设计了一种二阶段的优良位形采样器，可高效生成有效的起始或终止位形。实验结果表明该优良位形采样器能快速获取优良的位形，并提高规划算法的效率和路径的质量。本研究利用深度学习技术获取规划算法所需的起吊和就位位形，从另一个角度对起重机器人运动规划进行了尝试，推动了自主吊装的发展，为起重机器人智能化应用奠定基础。