毕业生知识库

随着机器人技术和人工智能的发展，越来越多的服务机器人被应用于商场、车站、机场、银行等公共场所，为人们提供各种服务，如导航、清洁、物品搬运等。对服务机器人而言，安全、高效地自主导航是最关键的能力之一，但现有的机器人导航技术还远未达到自主适应各种与人类交互场景的智能水平。因此，人-机共融场景中的机器人社交导航技术是目前的研究重点，具体而言：服务机器人不仅要能够准确理解人-机-环境三者之间的交互关系，还需要具备智能的运动规划能力，以实现安全导航的同时，遵守人类的社交规范。研究面向服务机器人的社交导航技术，能够进一步提高服务机器人与人和环境共融的能力，有利于推动服务机器人在实际生活中的普及和应用。

近年来，机器人社交导航技术取得了一些进展。但是仍然面临以下三个方面的挑战：（1）行人之间的交互关系会对机器人的导航决策产生间接影响，然而这些交互关系是复杂且难以观测的，传统的交互建模方式往往无法很好地处理行人之间的交互关系。（2）行人的行为受到多种因素的影响而表现出很强的随机性，难以通过简单的规则或模型来准确地预测其未来的移动轨迹。同时，机器人的运动规划需要具备较强的鲁棒性才能应对各种不确定性情况。（3）在人-机-环境共融的场景中，环境结构、行人的行为模式以及机器人的动作策略之间存在着复杂的交互关系，难以准确表征，这对机器人的交互关系建模和场景理解能力提出了更高的要求。同时，如何利用人-机-环境之间的交互信息来进行机器人的运动规划，仍然需要进一步研究和探索。本文针对面向服务机器人的社交导航方法开展研究，主要工作和创新点归纳如下：

（1）针对社交导航任务中人类交互意图难以观测的问题，提出一种基于卷积社交池化网络的机器人运动规划方法，实现了多人交互环境下机器人高效自主运动。首先，建立了社交导航问题的强化学习框架，设计了相应的状态空间、动作空间以及奖励函数；其次，构建了卷积社交池化网络，实现了对行人之间交互关系以及机器人与行人之间交互关系的特征提取；然后，构建了基于Actor-Critic强化学习的策略模型，通过考虑社交场景中的行人交互关系进行机器人动作决策。最后，通过仿真环境和真实场景中的大量实验验证了所提方法的有效性。

（2）针对社交导航任务中人类行为的不确定性高、难以准确预测的问题，提出一种基于行人多模式运动预测的机器人运动规划方法，提高了多行人动态环境下机器人运动规划的鲁棒性。首先，构建了时空交互特征提取模块以提取行人之间交互的时空相关性；然后，基于状态空间模型构建具有随机组件的行人轨迹预测模型，并且将机器人规划的动作嵌入到行人状态的预测过程中，以考虑行人可能的响应行为，实现对每个行人多样化潜在行为的推断；接着，提出了一个基于群体搜索的机器人动作决策算法，利用预测出的行人多条潜在轨迹优化机器人动作序列，实现机器人高效、鲁棒运动的同时遵守社交规范；最后，在仿真环境和真实场景进行了大量实验，结果表明所提方法在导航成功率、效率等指标上优于目前最先进的方法。

（3）针对环境-行人-机器人三者之间交互关系难以准确表征的问题，提出一种基于异质关系图学习的机器人运动规划方法，提高了机器人在人-机-环境共融场景下动作决策的能力。首先，提出了一种环境信息矢量化的特征表征方式，将障碍信息转化为图结构数据，并利用图神经网络提取其结构化特征；其次，构建了一个基于异质关系图网络的多模态交互特征融合模块，实现了对行人与行人、行人与障碍之间的异质交互特征的融合；然后，设计了一个机器人策略-价值模型，通过自动分配机器人-障碍交互特征、机器人-行人交互特征以及机器人目标特征的权重，实现了在不同场景下机器人策略的自适应调整；接着，构建了基于蒙特卡洛树搜索的动作决策框架，实现了机器人在复杂环境下的自主导航；最后，通过与最先进的算法进行对比实验，验证了所提方法的有效性。

With the development of robotics technology and artificial intelligence, more and more service robots are being deployed in public places such as shopping malls, stations, airports, and banks, providing various services to people, such as navigation, cleaning, and luggage handling. For service robots, safe and efficient autonomous navigation is one of the most critical capabilities. However, the existing robot navigation technology has not yet reached the level of intelligence required for autonomously adapting to various human interaction scenarios. Therefore, social navigation techniques for robots in human-robot collaborative scenarios are a current research focus. Specifically, service robots not only need to accurately understand the human-robot-environment interactions, but also require intelligent motion planning capabilities to achieve safe navigation while respecting social norms. Research on social navigation technology for service robots can further improve the ability of service robots to integrate with human and the environment, and is conducive to promoting the popularity and application of service robots in real life.

In recent years, there have been some advances in robot social navigation technology. However, it still faces the following three challenges: (1) The interactions among pedestrians indirectly affect the navigation decisions of robots. However, these interactions are complex and difficult to observe. Traditional interaction modeling methods can not handle the inter-human interactions effectively. (2) Human behaviors are influenced by multiple factors and exhibit strong stochasticity, making it challenging to accurately predict their future trajectories by simple rules or models. Meanwhile, robot's motion planner requires strong robustness to deal with various uncertainties. (3) In human-robot-environment coexistence scenarios, there are complex interactions between environment structures, pedestrian behavioural patterns, and robot action policies, which are difficult to be accurately characterised. This imposes higher demands on the robot's capability for interaction modeling and scene understanding. Moreover, how to utilize human-robot-environment interactions to inform robot motion planning still needs further research and exploration. This work focuses on the research of social navigation methods for service robots. The main contributions of this paper are summarized as follows:

(1) To address the difficulty of observing inter-human interactions in social navigation tasks, a robot motion planning method based on convolutional social pooling network is proposed, which enables the robot to navigate autonomously and efficiently in a multi-human environment. Firstly, a reinforcement learning framework for the social navigation task is established, and the corresponding state space, action space, and reward function are designed. Then, a convolutional social pooling network is constructed to extract features of the inter-human interactions and interactions between the robot and human. Finally, a policy model based on Actor-Critic reinforcement learning is developed, which makes action decisions for the robot by considering inter-human interactions in social scenarios. Finally, the effectiveness of the proposed method is verified through extensive experiments in both simulation environments and real-world scenarios.

(2) To address the high uncertainty of human behaviors and the difficulty of accurately predicting their future behaviors in social navigation tasks, a robot motion planning method based on multi-modal pedestrian motion prediction is proposed, which improves the robustness of robot motion planning in dynamic multi-human environments. Firstly, a spatio-temporal interaction feature extraction module is constructed to extract the spatio-temporal correlation of inter-human interactions. Secondly, a pedestrian trajectory prediction model with a stochastic component is constructed based on the state-space model, and the robot's motion plans are incorporated into the prediction process of the pedestrians' states in order to take into account the possible response behaviors from pedestrians. The prediction model is capable of reasoning about diverse potential behaviors for each pedestrian. Thirdly, a robot action decision-making algorithm based on population search is proposed, which takes into account the multiple possible predictions and optimizes robot's action sequences to align with social norms, ultimately providing more robust actions. Finally, extensive experiments are conducted in both simulation environments and real-world scenarios. The results demonstrate that he proposed method outperforms the current state-of-the-art methods in terms of the success rate of navigation, efficiency, and other metrics.

(3) To address the challenge of accurately characterizing the interaction relationships between the environment, human, and robots, a robot motion planning method based on heterogeneous relation graph learning is proposed, which improves the robot's ability to make action decisions in human-robot-environment coexistence scenarios. Firstly, our approach represents the feature of environment information in a vectorized form, where the obstacle information is transformed into graph-structured data and then its structured features are extracted by graph neural networks. Secondly, a multi-modal interaction feature fusion module based on heterogeneous relation graph network is constructed, which is capable of fusing inter-human interactions and human-obstacle interactions. Thirdly, a robot policy-value model was designed, which is capable of adapting to different scenarios by automatically adjusting the importance of robot-obstacle interaction features, robot-human interaction features, and goal features. Fourthly, an action decision-making framework based on Monte Carlo Tree Search is built, which enables robot to navigate autonomously in complex environments. Finally, the effectiveness of the proposed method is verified by comparison experiments with other state-of-the-art algorithms.