欠明确表达任务下基于非受限交互的机器人模仿学习方法研究

With regard to the requirements of interactive and manipulative service robots, this research project will study the theory and method of robot imitation learning based on non-restricted human-robot Interaction for goal-directed and under-specified tasks. The proposed study will consist of the following five aspects. ①In order to deal with the problem of high-dimensional feature space in imitation learning, a dimensionality reduction method based on Locality Preserving Projection Gaussian Process Latent Variable Model is proposed. ②Using the RGB-D information of depth cameras, a study on the representation and learning method of human demonstration behaviors is performed, in which the problems of noisy observation and large-scale sparse structure are to be solved. ③The grasping configuration of irregular objects is difficult to obtained using learning-by-demonstration method only. With the consideration of the under-specified nature of the human demonstrated tasks, a new grasp learning methodology will be proposed that combines imitation learning and try-and-error learning strategies. As a result of model learning and inference, a generalized grasping action that satisfies both task and force constrains can be created. ④A unified human-robot heterogeneous mapping mechanism will be proposed by using semantic technology and robot knowledge representation framework. Thus the mapping between abstracted concepts and robot plans can be achieved. ⑤Experiments on the typical manipulation tasks of service robots, which mainly include a set of pick-place-move-and-rotate primitives, will be conducted using the experimental platform. Hence experimental verification will make the theory and method more perfect. The proposed work will provide a systematic theory and key implementation techniques of service robots interactive learning systems.

本课题拟针对作业型交互式服务机器人需求，研究欠明确表达任务下基于非受限交互的机器人模仿学习方法，包括：①研究提出小样本高维数据的保局高斯过程潜变量模型，解决模仿学习观测样本高维冗余特征降维方法；②利用深度视觉传感器RGB-D信息，研究提出非受限交互下基于物品可供性的演示操作行为表征与学习方法，解决传感器含噪和样本大规模稀疏带来的问题；③针对欠明确表达的灵巧手不规则物体抓取任务，研究提出融合模仿与试错两种学习机制的抓取推理与行为泛化理论方法，实现满足任务-力双重约束的机器人抓取动作重构；④引入语义技术和机器人知识描述框架，研究提出与操作对象无关的统一人机异构映射与执行机制，建立从抽象概念到机器人可执行计划之间的映射关系；⑤结合含抓取-放置-平移-旋转等基元动作序列组成的服务机器人典型操作任务应用，通过实验研究验证并完善相关理论，形成一套新型服务机器人交互学习应用系统设计方法。

针对作业型服务机器人基于视频的联动/非联动两类典型操作任务示教行为示范学习，深入研究了目标导向型（Goal-directed）欠明确（Under-Specified）表达任务的机器人模仿学习理论与方法，具体研究取得了以下成果：①研究了基于GPLVM的高维手爪姿态特征降维算法，实现了20维抓取姿态特征降维；②在非受限交互前提下，引入物品可供性思路，构建了基于CRF的操作示范行为中人与物时空关系建模，在此基础上研究了视频动作序列分割与识别方法；③针对机器人可泛化的操作行为复现，将动作序列分解为动作基元的序列组合；针对抓取动作基元的核心问题——抓取姿态模仿学习，提出并不直接模仿人手姿态、而是学习示范中蕴含的与任务相关的机器人抓取位姿约束，从而赋予机器人从示范中学习任务相关的抓取约束和对特定物体实例推理其任务相关抓取姿态的能力；针对轨迹类动作基元，提出任务参数化DMP模型学习动作的多演示轨迹，从而提高轨迹对环境变化和轨迹参数变化的泛化能力；④为确保所学技能在机器人重现时的可泛化性能，采用任务学习而非轨迹克隆思路，建立了人机映射与执行机制，将任务转换为机器人可执行的语义化计划指令，并借助PDDL来推理示范中缺失的抓取细节，将机器人计划指令通过具体待操作、待抓取/放置的位姿、轨迹等参数来实例化，从而实验了机器人动作复现。⑤在UR5机器人平台上，选取了含抓取-放置-平移-旋转等基元动作序列组成的典型作业任务，开展了实验研究与完善，验证了所提方法的有效性。