标题： 从感觉运动动态功能连接中无监督发现行为基元 显示英文标题

标题： Unsupervised Discovery of Behavioral Primitives from Sensorimotor Dynamic Functional Connectivity

摘要： 动物和机器人运动产生高维的运动和感觉信息流。 想象一个新生儿的大脑或一个婴儿人形机器人的控制器试图理解未经处理的感觉运动时间序列。 在这里，我们提出一个框架，用于研究机器人代理的多模态感觉信号之间的动态功能连接，以揭示潜在的结构。 使用瞬时互信息，我们捕捉本体感觉、触觉和视觉信号之间随时间变化的功能连接（FC），揭示感觉运动关系。 使用无限关系模型，我们识别了感觉运动模块及其演变的连接性。 为了进一步解释这些动态交互，我们采用了非负矩阵分解，将连接模式分解为加法因子及其对应的时间系数。 这些因子可以被认为是代理的运动基元或运动协同作用，代理可以利用它们来理解其感觉运动空间，并在以后用于行为选择。 在未来，该方法可以部署在机器人学习中，也可以用于人类运动轨迹或脑信号的分析。 显示英文摘要

摘要： The movements of both animals and robots give rise to streams of high-dimensional motor and sensory information. Imagine the brain of a newborn or the controller of a baby humanoid robot trying to make sense of unprocessed sensorimotor time series. Here, we present a framework for studying the dynamic functional connectivity between the multimodal sensory signals of a robotic agent to uncover an underlying structure. Using instantaneous mutual information, we capture the time-varying functional connectivity (FC) between proprioceptive, tactile, and visual signals, revealing the sensorimotor relationships. Using an infinite relational model, we identified sensorimotor modules and their evolving connectivity. To further interpret these dynamic interactions, we employed non-negative matrix factorization, which decomposed the connectivity patterns into additive factors and their corresponding temporal coefficients. These factors can be considered the agent's motion primitives or movement synergies that the agent can use to make sense of its sensorimotor space and later for behavior selection. In the future, the method can be deployed in robot learning as well as in the analysis of human movement trajectories or brain signals.