中国科学院自动化研究所机构知识库

海洋世界深邃而神秘，探索海洋是人类永恒的主题。近年来，随着智能制造和人工智能技术的发展，水中作业机器人逐渐成为人类进一步开发和利用海洋资源的重要利器。本文围绕水中智能作业机器人的视觉感知与运动控制这一主题，以实现水中作业机器人“看得精准”、“游得平稳”、“作业聪慧”为目标，从水下和水面两个空间维度，在水下双目相机标定方法、水下轨迹跟踪控制、基于视觉的水面智能清洁机器人的运动控制与抓取决策、面向任务的水面综合路径规划体系等方面展开了深入研究，主要内容如下：

一、为实现水中作业机器人“看得精准”，提出了一种基于折射校正的水下双目相机标定方法。首先，构建了水下相机折射模型，清晰地呈现了折射条件下的水下相机成像过程。其次，提出了类三角化水下目标点的定位方法，并确定水下双目相机的待标定参数。再次，利用标定板角点的相对位置不变性，建立了距离不变性、垂直不变性、平行不变性三个优化目标，从而将水下双目测量系统的标定问题转化为多目标优化问题，并用NSGA-II算法进行求解。最后，经标定的双目相机，在水下目标位置与物体尺寸测量上均有良好的表现，并成功应用于水下自主垃圾清理作业之中。

二、为确保水中作业机器人“游得平稳”，针对水下航行器控制中存在的三个主要问题，即欠驱动特性、速度约束、集总干扰，提出了一种融合模型预测与扩展状态观测的水下三维轨迹控制方法。在模型预测调配器方面，推导出了欠驱动形式的运动学跟踪误差模型方程，设计满足速度约束的模型预测代价函数，并基于此生成轨迹跟踪的参考速度；同时，为消除动力学耦合造成的跟踪静态误差，提出了偏航角补偿机制以提升跟踪精度。在动力学控制器设计方面，设计了扩展状态观测器以观测集总干扰与不易测量的速度状态；同时，基于扩展状态观测器，设计相应的动力学控制律，实现了精准的三维轨迹跟踪控制。通过仿真分析和水池实验对所提方法进行了性能评估，验证了该轨迹跟踪控制方法的有效性。

三、研制了一种能够同时实现“看得精准”、“游得平稳”、“作业聪慧”的基于视觉的水面智能清洁机器人系统，具备自主巡航与垃圾检测、视觉导航与目标接近、垃圾抓取与自动收集三项功能。在漂浮垃圾检测方面，搭建了基于YOLOv3的目标检测网络，并构建漂浮垃圾数据集，实现了实时精准的检测。在视觉导航控制方面，设计了基于滑模的视觉导航控制器以提升机器人的抗干扰能力，从而保证机器人平稳接近目标。在目标抓取方面，分析了较难抓取的细长物体在水流环境中的平衡特性，基于此提出了预判抓取策略，成功实现垃圾抓取。最终，室内与野外测试结果表明，该系统具有自主清洁漂浮垃圾的能力和潜在的应用推广价值。

四、依据任务特点实现自主规划与决策是“作业聪慧”的重要体现，为此构建了面向水面自主作业任务的综合路径规划体系。该体系包括任务分配、初规划、重规划三个环节。具体地，将任务分配问题转化为旅行商问题，使用基于FM*的距离度量替代了欧几里得距离度量，以获得障碍物条件下的最优遍历序列。在初规划环节，利用高斯滤波器来降低生成路径的转弯难度以满足机器人机动性约束，同时保持与障碍物的安全距离。针对动态水环境中重规划频繁发生的问题，设计了基于神经网络的重规划点生成器以权衡路径代价与计算代价。仿真分析与虚拟障碍物实验结果表明，该综合路径规划体系可为水面自主清洁作业提供有效的路径信息与高效实时的重规划决策。

The ocean world is deep and mysterious, and exploring the ocean is the eternal theme of mankind. In recent years, with the development of intelligent manufacturing and artificial intelligence technologies, underwater robots have gradually become an important weapon for humans to further develop and utilize marine resources. This article focuses on the theme of visual perception and motion control of underwater intelligent operating robots, with the goal of achieving underwater robots “accurate observing”, “smooth swimming”, and “smart operating”. From the two spatial dimensions of underwater and water surface, in-depth research has been carried out on underwater binocular camera calibration methods, underwater trajectory tracking control, vision-based motion control and grasping decision-making of intelligent water surface cleaning robots, and task-oriented water surface comprehensive path planning system. The main contents are as follows:

Firstly, in order to realize the “accurate observing” of aquatic robots, a calibration method for underwater binocular cameras based on refractional correction is proposed. For a clear presentation of the imaging process of the underwater camera under the condition of refraction, the refraction model of the underwater camera is constructed. Subsequently, the akin triangulation locating method is proposed for underwater target, and the parameters to be calibrated for the underwater binocular camera are determined. Furthermore, using the relative position invariance of the corner points of the calibration plate, three optimization goals of distance invariance, vertical invariance, and parallel invariance are established, so as to convert the calibration problem of the underwater binocular measurement system into a multi-objective optimization problem that is solved by NSGA-II algorithm. The calibrated binocular camera has good performance in the measurement of underwater target position and object size, and has been successfully applied in underwater autonomous garbage cleaning operations.

Secondly, in order to ensure the “smooth swimming” of the underwater working robot, a fusion model predictive control and extended state are proposed for the three main problems of the underwater vehicle control project in the control of underwater vehicles, namely underactuated characteristics, speed constraints, and lumped disturbances. In the aspect of the model predictive governor, the underactuated kinematics tracking error model is derived, and the model predictive cost function that satisfies the speed constraint is designed. Based on this, the referenced velocities of the trajectory tracking are generated. Meanwhile, to eliminate the static tracking error caused by the dynamic coupling, a heading angle compensation mechanism is proposed to improve tracking accuracy. In terms of dynamic controller design, the ESO is designed to observe the lumped disturbance vector and the difficult-to-measure speed states. At the same time, based on the extended state observer, the corresponding dynamic control law is designed to achieve precise three-dimensional trajectory tracking control. The performance of the proposed method is evaluated through simulation analysis and experimental verification, and the effectiveness of the trajectory tracking control method is verified.

Thirdly, in order to enable underwater robots to simultaneously realize “accurate observing”, “smooth swimming”, and “smart operating”, a vision-based intelligent water surface cleaning robot system has been developed to realize autonomous cruise and garbage detection, visual navigation, and garbage grasping and collection. In terms of floating garbage detection, a target detection network based on YOLOv3 is built, and a floating garbage data set is constructed to achieve real-time and accurate detection. In terms of visual navigation control, a sliding-mode-based visual navigation controller is designed to improve the anti-interference ability of the robot, so as to ensure that the robot approaches the target smoothly. In terms of target grasping, the balance characteristics of slender objects that are difficult to grab in a water flow environment are analyzed. Based on this, an early decision grasping strategy is proposed to successfully achieve garbage grasping. In the end, results of indoor and field tests show that the system has the ability to clean floating garbage autonomously and has wide application prospect.

Finally, the realization of autonomous planning and decision-making based on the characteristics of tasks is an important manifestation of “smart operating”. For this purpose, a comprehensive path planning system for autonomous surface operations tasks has been constructed. The system includes three links: task allocation, preliminary planning, and re-planning. Specifically, the task assignment problem is transformed into the travelling salesman problem, and the Euclidean distance metric is replaced by the FM*-based distance metric to obtain the optimal traversal sequence under obstacle conditions. In the initial planning stage, Gaussian filters are used to reduce the turning difficulty of the generated path to meet the robot's mobility constraints, while maintaining a safe distance from obstacles. Aiming at the frequent re-planning problems in the dynamic water environment, a neural network-based re-planning point generator is designed to strike compromise between the distance cost and the calculation cost. Results of simulation case analysis and virtual obstacle experiment show that the integrated path planning system can provide effective path information and efficient real-time re-planning decisions for autonomous water surface cleaning operations.