智能机器人的主要工作流程包含感知、决策、规划和控制。随着机器人智能化的发展，在感知层面，主动感知的相关研究得到了关注。本文主要研究多自由度机器人主动感知关键技术，这里的多自由度同时包含了机械臂以及移动机器人，这两者与传感器共同构成了本文的硬件主体。而本文的软件主体是主动感知与防碰撞两项技术，分别属于机器人的感知与执行两大步骤。目前的多自由度机器人主动感知技术在复杂场景、复杂任务的情况下仍存在实时性无法满足、安全性无法确保和准确率不够等问题。本文从实际的机器人应用场景出发，尝试从视点规划、运动规划、导航避障和感知处理等方面去解决这些问题，主要的创新与贡献包括：1.本文提出了同时包含排序与评估的双分支特征交互融合的最佳视点计算模型来计算最优的采集视点，提高了最佳视点的计算准确率；提出了根据机械臂的运动学特点来设计的基于参数空间分解的组合快速碰撞检测分类器，用以提高碰撞检测的准确率和速度，进而加速机械臂运动规划过程并确保实时性；构建了一套将计算复杂度从三维降到二维的平面内卷积平面间积分的高效可达性图生成方式，并进一步来计算机械臂最优的基底安装位置以及重分配视点采样的密度，提高的可达性图计算与运动规划的效率。通过以上三方面的改进，本文提高了多自由度机械臂的主动视觉与防碰撞性能。2.本文设计了一种低成本的基于传感器融合与自状态注意力机制的移动机器人防碰撞模型，针对具有一定高度的机器人给出了一套低成本的解决方案，并提高了深度强化学习模型的效率。本文还设计了一种双头Transformer深度强化学习防碰撞模型，进一步提高了模型在复杂场景下的适应能力，提高了机器人的避障效率。通过以上两方面的改进，本文提高了由移动机器人与机械臂搭建的主动感知机器人的性能。3.本文提出了一项基于点云补全和关键点配准的6D位姿估计后处理方法，设计了一种基于融合特征、多目标优化的点云补全网络和颜色支持的迭代关键点算法。该后处理方法在当前最佳方法的基础上进一步提高了6D位姿估计的准确率。基于主动视觉的不同的任务会产生不同的感知处理、信息使用算法，该创新点以6D位姿估计任务这一机器人主动感知的核心应用之一为研究对象，描述感知处理与信息使用与具体任务相结合的流程，进一步拓展了本文的应用范围。

The main pipeline of intelligent robots includes perception, decision-making, planning, and control. With the development of robot intelligence, researchers have gradually begun to pay attention to active perception at the perception level. This paper mainly does research the key technologies of the active perception of multiple degrees-of-freedom robots. This type of robot contains robotic arms and mobile robots. The robotic arm, mobile robot, and sensors compose the hardware of this paper. The main technologies of the software in this paper are active perception and collision avoidance, which correspond to the perception and execution of the robot. The current active vision technology for multi-degree of freedom robots still has problems such as insufficient real-time performance, insufficient safety, and insufficient accuracy. Starting from robot application, this paper attempts to solve these problems from view planning, motion planning, navigation, and perception processing. The main innovation and contributions of this paper include:1. This paper has designed the next best viewpoint computing model of double branch neural network that includes both ranking and evaluation to calculate the optimal collection viewpoint, improving the calculation accuracy of the best viewpoint, proposes a configuration space decomposition for Learning-based collision checking in high-DOF robots according to the kinematics characteristics of the manipulator, improving the accuracy and speed of collision detection, and constructs an intra-plane convolution of inter-plane integration method to easily get the reachability map which reduces the computational space from 3D to 2D. The reachability map is used to find the best robotic arm base position and a new sampling map for viewpoint, greatly improving the computational efficiency. Through the improvement of the above three points, this paper improves the active vision and anti-collision performance of the multiple degrees-of-freedom robotic arm.2. This paper designs a low-cost solution for a mobile robot with sensor fusion and self state attention mechanism. Besides, this paper also designed a double-headed transformer to deeply strengthen the reinforcement learning model, which further improves the adaptation ability of the model in complex scenes and improves its efficiency. Through the improvement of the above two points, this paper improved the performance of active perception for robots built by mobile robots and robotic arms.3. This paper proposes a 6D pose post-processing method based on point cloud completion and key points registration. It designs a point cloud completion algorithm based on fusion features and multi-target optimization networks and color support key point registration. The post-processing method further improves the accuracy of the 6D position estimation on the basis of the SOTA, and verifies the effectiveness of our method. Different tasks based on active vision will generate different perception processing algorithms. The innovation is based on one of the core applications of the active perception of the robot active perception of the task of the 6D pose estimation and it further expands the application scope of this paper.