毕业生知识库

机器人任务序列问题（Robotic Task Sequencing Problem，RTSP）是机器人在处理多任务时必须解决的一个关键问题，需要在涉及空间约束、资源约束和运动学约束等多种约束条件的情况下，同时优化相互耦合的任务执行序列以及机器人运动轨迹，问题的复杂度高且难以解决。当涉及多机器人系统时，还需要考虑多机器人协调作业和避免碰撞等问题，进一步增加了问题求解的复杂性。现有方法大多研究结构化场景下的RTSP问题，未考虑半结构化场景下机器人的应用且面对大规模任务时求解效率与求解质量差，严重限制了机器人的应用范围。针对上述问题，本文围绕着半结构化场景中的机器人多任务轨迹规划问题，从单机器人RTSP求解、多机器人任务的RTSP求解、面向半结构化场景的RTSP实时求解以及半结构化场景位姿计算等方面展开研究。本文的主要工作和贡献如下：

1.      针对RTSP求解复杂度高的问题，提出一种基于解耦策略的高效求解方法，利用机器人任务空间与关节空间之间的结构特点和属性差异，将RTSP问题分解为任务执行序列优化与任务转移轨迹计算并分别在任务空间与关节空间进行求解，在保证求解质量的同时可显著提升了计算效率；利用碰撞检测和启发式关节构型筛选机制减少机器人的运动学冗余，可在关节空间中保证任务之间转移运动的轨迹平滑性。实验结果表明：所提方法显著提升了RTSP问题的求解效率，生成的机器人运动轨迹安全有效，可以提高机器人应对多任务的执行效率。

2.      针对多机器人任务RTSP问题的高复杂性，提出了一种考虑资源约束并将整体问题进行解耦求解的方法，通过对问题进行合理简化，依次求解多机器人系统中任务调度、资源分配、关节构型分配和运动协调等子问题，以优化多机器人系统的任务执行效率。结合任务调度及资源分配的特点，将其转化成最短路径问题并应用图优化技术求解，以提升问题的求解效率。考虑其他机器人对关节构型选择的影响，借助第1项工作所提关节构型分配方法解决多机器人的最优关节构型问题，并采用集中式运动规划来求解多机器人之间的无碰撞运动轨迹。实验结果表明：所提方法能够高效求解多机器人任务的RTSP问题并保证求解质量，所生成的机器人运动轨迹平滑性好，能满足实际应用的需求。

3.      针对半结构化场景RTSP求解的高实时性要求，提出一种基于RTSP聚类求解和数据驱动的半结构化场景轨迹复用策略。首先通过对任务点进行聚类构建聚类旅行商问题提升了第1项工作中所提方法的求解速度，同时采用并行处理机制来规划和执行机器人的运动轨迹；然后针对不可复用的关节构型分配提出了基于超网络架构和高斯混合模型的网络，该网络利用机器人运动学模型的链式结构和历史关节构型分配结果，保证了关节构型分配的合理性与最优性；最后结合历史规划复用及所提关节构型分配网络，提出了用于实时求解半结构化场景中动态RTSP问题的方法，将大规模RTSP问题的求解时间减少至0.1秒以内，并保证了求解质量以及机器人运动轨迹的平滑性。

4.      针对半结构化场景RTSP求解所需的物体位姿计算问题，提出了一种对噪声有较强鲁棒性的对应预测网络，利用基于稀疏卷积的特征编码器对输入的点云进行局部特征提取和下采样，以提升对噪声的鲁棒性和运算速度；通过自注意力和交叉注意力机制聚合点云内部的全局信息，促进点云间的信息交互，并更新提取得到的局部特征；同时通过正弦位置编码将点云的空间信息融入其中；解码器根据所提取的特征，预测下采样点在另一点云坐标系中的位置及其在重叠区域的置信度，以建立精确的点对应关系；通过加权奇异值分解，从点对应关系中计算变换矩阵，实现精确的点云配准。所提方法在公开数据集3DMatch和3DLoMatch上与最先进的方法相比，展示了优异的配准性能，并在两个噪声数据集上表现出最佳的抗噪声鲁棒性。

The Robotic Task Sequencing Problem (RTSP) is a pivotal challenge faced by robots when handling multiple tasks, subject to various constraints including spatial, resource, and kinematic constraints. It poses a formidable challenge to simultaneously optimize the sequencing of task execution and robot motion trajectories. While existing methodologies predominantly focus on addressing RTSP in structured environments, they often overlook the deployment of robots in semi-structured environments and their efficiency and solution quality diminish significantly when confronted with large-scale task scenarios, severely constraining the applicability of robots. In light of the above issues, this dissertation revolves around the multi-task trajectory planning problem of robots in semi-structured environments, and conducts research on single-robot RTSP solving, multi-robot task RTSP solving, real-time RTSP solving for semi-structured environments, and pose calculation for semi-structured environments. The main work and contributions of this dissertation are as follows:

1.      To address the high computational complexity of RTSP solving, a highly efficient solving method based on decoupling strategy is proposed. By utilizing the structural characteristics and attribute differences between robot task space and joint space, the RTSP problem is decomposed into optimizing task execution sequences and calculating task transfer trajectories, which are solved separately in task space and joint space respectively. This method significantly improves computational efficiency while ensuring solution quality. Additionally, collision detection and heuristic joint configuration selection mechanisms are utilized to reduce the kinematic redundancy of robots, ensuring the smoothness of task transfer trajectories in joint space. Experimental results demonstrate that the proposed method significantly enhances the efficiency of RTSP problem solving, generating safe and effective robot motion trajectories that can improve the efficiency of robots in handling multiple tasks.

2.      To address the high complexity of the multi-robot task RTSP problem, a method considering resource constraints and decoupling the overall problem for solving is proposed. By reasonably simplifying the problem, sub-problems such as task scheduling, resource allocation, joint configuration assignment, and motion coordination in multi-robot systems are solved sequentially to optimize the task execution efficiency of the multi-robot system. By transforming task scheduling and resource allocation into shortest path problems and applying graph optimization techniques to solve them, the efficiency of problem solving is enhanced. Considering the influence of other robots on joint configuration selection, the optimal joint configuration problem for multiple robots is solved using the joint configuration assignment method proposed in the first step, and centralized motion planning is used to solve collision-free motion trajectories between multiple robots. Experimental results demonstrate that the proposed method can efficiently solve the RTSP problem of multi-robot tasks while ensuring solution quality, generating smooth robot motion trajectories that meet practical application requirements.

3.      In response to the high real-time requirements of solving RTSP in semi-structured environments, a trajectory reuse strategy based on RTSP clustering solving and data-driven approaches is proposed. Firstly, by clustering task points to construct a clustering traveling salesman problem, the speed of the method proposed in the first step is improved, and a parallel processing mechanism is used to plan and execute robot motion trajectories. Next, an architecture based on hypernetworks and Gaussian mixture models is proposed for non-reusable joint configuration assignments, utilizing the chain structure of the robot kinematic model and historical joint configuration assignment results to ensure the rationality and optimality of joint configuration allocation. Finally, combining historical planning reuse and the proposed joint configuration assignment network, a method for real-time solving dynamic RTSP problems in semi-structured environments is presented, reducing the solving time of large-scale RTSP problems to within 0.1 seconds while ensuring solution quality and smooth robot motion trajectories.

4.      In response to the object pose calculation required for solving RTSP in semi-structured environments, a correspondences prediction network with strong robustness to noise is proposed. By utilizing a feature encoder based on sparse convolution to extract local features and downsample the input point cloud, robustness to noise and computational speed are improved. The network aggregates global information within the point cloud through self-attention and cross-attention mechanisms to promote information interaction between point clouds and update the extracted local features. Additionally, spatial information of the point cloud is integrated using sine positional encoding. The decoder predicts the positions of downsampled points in another point cloud coordinate system and their confidence in the overlapping areas based on the extracted features to establish precise point correspondences. Weighted singular value decomposition is used to calculate transformation matrices from point correspondences, enabling accurate point cloud registration. The proposed method demonstrates excellent registration performance compared to state-of-the-art methods on public datasets 3DMatch and 3DLoMatch, showing superior robustness to noise on two noise datasets.