Robot arm and hand manipulation using vision-guided motion control directed by human gestures

点击翻译按钮获取中文摘要

This research will focus on the geometric specification of robot (arm and hand) manipulation tasks in camera (and other sensory) coordinate frames. We propose to develop Human-Robot Interfaces (HRI) based on pointing and gestures. The gestures will be interpreted through computer vision and robot motions will be performed through visual servoing, and force servoing. The main objective is to solve general robot arm and hand manipulation tasks interactively with humans. Robotics research in both engineering and artificial intelligence has focused on autonomy. While autonomy works well in specific situations (such as factory automation), despite decades of research, robotics cannot solve human arm and hand manipulation tasks in everyday unstructured environments. However, in many cases autonomy may not be the goal. Humans naturally interact and collaborate on tasks. A human can easily guide another human how to perform complex tasks. Much of this interaction is non-verbal. For robots to work with humans they need to take visual cues from the human, and be easy to interact with, rather than working autonomously. Hence, with good HRI, humans could relatively effortlessly supervise tasks that are otherwise too complex for robots. We propose an integrated systems and experiment-driven approach to this research. In particular we will research: 1. Task specification. We will develop the so called "task-function" to handle a combination of camera image-, 3D world- and force sensor coordinates. 2. Video tracking of image features. We will research how to improve robustness of registration-based tracking. 3. Hybrid Visual Servoing. We will extend image-coordinate visual servoing to also work with 3D and force sensory feedback, while requiring a minimum of a-priori calibration. 4. HRI. We will research gesture based human-robot interaction, including what geometric specifications (2D image, 3D, force contacts) are most natural for a human in order to specify complete, provable 3D alignments; and then to sequence into full manipulation tasks. We will develop benchmarks for everyday human manipulation tasks, and study our task specification interfaces on these (e.g. task completion times and cognitive load.) The goal of our research will be a better understanding of visual and sensory-space motion control, both from a robotics and human interface perspective. Practical outcomes include integration of different tracking and sensory-motor control methods into modular ROS (Robot Operating System) software libraries, and test results of this software by humans on many different manipulation tasks.