Model predictive motion planning for robot-assisted observation and recording of humanactivities

人类活动的真实数据在许多领域都是必不可少的，但同时获取起来又困难和昂贵。例如，数据用于医学、航空航天领域的文档目的，也用于行为学领域的人机交互研究。在深度学习或观察学习等机器学习中，数据的数量和质量决定了新方法的性能。机器人对人类活动的观察和记录是一种很有前途的方式，可以经济地获取稀有和独家数据。模型预测控制是目前机器人在避碰、时间最优和路径最优等方面进行实时运动规划的最新技术。因此，该研究项目致力于建模预测运动规划，以便通过机器人引导的眼控相机无缝记录人类活动和相关环境变化。目标是在不干扰人类的情况下，从接近活动的不同角度记录，从而最大限度地提高整个摄像机运动中积累的信息增益。随着模型预测控制在机器人技术中的成功应用，这一目标带来了新的挑战，既包括在人类运动不确定性下的碰撞避免，也包括系统地开发任务特定成本函数，以评估相对于更高层次的总体目标的整个时空运动的成功。这些方法将在一个示例应用的背景下进行详细和集体的评估，其中机器人操纵器将通过从观察中学习来再现观察到的活动。

Authentic data of human activities are essential in many fields but at the same time difficult and costly to acquire. Data is used, for example, for documentation purposes in medicine, aerospace, and also in ethology for research on human-robot interaction. In machine learning such as deep learning or learning from observation, the quantity and quality of data determine the performance of new approaches. Robotic observationand recording of human activities is a promising way to economically access rare and exclusive data. Nowadays, model predictive control is state of the art for real-time motion planning under the aspects of collision avoidance as well as time and path optimality for robotic manipulators. The research project is therefore dedicated to model predictive motion planning for a seamless recording of human activities and the associated changes to the environment by a robot-guided eye-in-hand camera. The goal is to maximize the information gain accumulated over the entire camera motion by recording from changing perspectives in close proximity to the activity without disturbing the human. With the successful application of model predictive control in robotics, new challenges arise following this objective, both, in terms of collision avoidance under uncertainties of human motion, and systematic development of task-specific cost functions that evaluate the success of the entire spatio-temporal motion with respect to a higher-level overall goal.The methods will be evaluated, both, in detail and collectively in the context of an exemplary application in which a robotic manipulator will reproduce the observed activity using learning from observation.