Real-time robotic tasks for cyber-physical avatars

摘要

Although modern robots can perform complex tasks using sophisticated algorithms that are specialized to a particular task and environment, creating robots capable of completing tasks in unstructured environments without human guidance (e.g., through teleoperation) remains a challenge. In this research, we present a framework to meet this challenge for a "cyberphysical avatar," which is defined to be a semi-autonomous robotic system that adjusts to an unstructured environment and performs physical tasks subject to critical timing constraints while under human supervision. This thesis first realizes a cyberphysical avatar that integrates three key technologies: (1) whole body-compliant control, (2) skill acquisition from machine learning (neuroevolution methods and deep learning), and (3) vision-based control through visual servoing. Body-compliant control is essential for operator safety because avatars perform cooperative tasks in close proximity to humans; machine learning enables "programming" avatars such that they can be used by non-experts for a large array of tasks, some unforeseen, in an unstructured environment; the visual servoing technique is indispensable for facilitating feedback control in human avatar interaction. This thesis proposes and demonstrates a systematically incremental approach to automating robotic tasks by decomposing a non-trivial task into stages, each of which may be automated by integrating the aforementioned techniques. We design and implement the controllers for two semi-autonomous robots that integrate three key techniques for grasping and pick-and-place tasks. While a general theory is beyond reach, we present a study on the tradeoffs between three design metrics for robotic task systems: (1) the amount of training effort for the robots to perform the task, (2) the time available to complete the task when the command is given, and (3) the quality of the result of the performed task. The tradeoff study in this design space uses the imprecise computation model as a framework to evaluate specific types of tasks: (1) grasping an unknown object and (2) placing the object in a target position. We demonstrate the generality of our integration methodology by applying it to two different robots, Dreamer and Hoppy. Our approach is evaluated by the performance of the robots in trading off between task completion time, training time and task completion success rate, in an environment similar to those in the recent Amazon Picking Challenge.