Vision-based supervisory control and predictive display for uncalibrated tele-robotics

我的研究涉及计算机视觉、机器人学和图形学。我特别感兴趣的是如何使用视觉信息和坐标框架来编码和交流空间信息、运动和交互。我从事的应用程序包括用于机器人运动控制的视觉伺服和用于操作员可视化的预测显示。这两者结合在一起，形成了远程机器人系统。他说，一个关键的想法是使用视觉(相机)空间，而不是传统工程中的欧几里得世界坐标。场景中人的指示性动作和手势可以被摄像机观看并在摄像机空间中被解释。那一幕。类似地，在远程机器人技术中，人类操作员沉浸在远程场景的可视化中，可以手势指示W.r.t.远程站点和本地站点之间的共享视觉空间。使用感官框架代替传统的世界坐标框架给任务和轨迹规划带来了新的问题。此外，在人类环境中运行的机器人面临的任务比目前在制造业中更多。在这种具有挑战性的情况下实现完全自主是困难的，而且可能并不总是可取的。相反，我们的目标是赋予机器人在人类指导下半自主地执行运动和小型子任务的能力，即所谓的监督控制。动作和任务是自下而上构建的，从简单的动作、伸展和精细操作到更复杂的子程序，例如挑选和放置。我们可以执行的任务包括灯泡交换子例程、设置餐桌和解决儿童形状分类谜题。这些例程目前是通过脚本构建的，但我们计划从观察人类中学习。视觉指导机器人的一般方法有许多应用，从高端，例如在外科辅助机器人中，其中人类外科医生不能中断手术来使用传统的键盘接口，到日常服务机器人，例如为残疾人和老年人服务。

My research is in the intersection of computer vision, robotics and graphics. In particular I'm interested in how to use visual information and coordinate frames to encode and communicate spatial information, motion, and interactions. Among applications I work on are visual servoing for robot motion control and predictive display for operator visualization. Both of these combine into tele-robotics systems.         A key idea is to work with visual (camera) spaces instead of Euclidean world coordinates as in conventional engineering. Deictic motions and gestures of a human in a scene can be viewed by cameras and interpreted in camera space w.r.t. that scene. Similarly, in tele-robotics, a human operator is immersed in a visualization of a remote scene and can gesture instructions w.r.t. the shared visual space between the remote and local sites. Using sensory frames instead of a conventional world coordinate frame brings up new issues in task and trajectory planning. Furthermore, robots operating in human settings face more varying tasks than currently in manufacturing. Achieving full autonomy in such challenging situations is difficult, and may not always be desirable. Instead our thrust is to endow the robot with the ability to carry out motions and small subtasks semi-autonomously, while under the guidance of a human, so called supervised control. The motions and tasks are built bottom up, from simple motions, reaching and fine manipulation to more complex subroutines, e.g. pick-and-place. Tasks we can implement include a light bulb exchange subroutine, setting a dinner table and solving a child's shape sorter puzzle. The routines are currently built by scripting, but we plan to look into learning from observing a human. A general way of visually instructing a robot has many applications from the high end e.g. in surgical assistive robotics, where the human surgeon can't interrupt the surgery to use a conventional keyboard interface, to everyday service robotics e.g. for the handicapped and elderly.