Basic research on detection of task errors in robotic manipulation

开发了一种灵巧操作系统，作为任务误差检测的试验台：该系统由一个多指手组成，每个手指上都装有力/扭矩传感器，三维实时立体视觉测量被操纵物体的姿态，以及一个实时手指控制器。开发了以下功能：(1)。通过指尖与被抓物体表面的滚动接触来操纵被抓物体，实现精细控制，伺服周期时间为10ms。3- d立体视觉每33ms测量一次被抓物体的姿态。提出了一种从力/扭矩传感器信号中分离内力和外力的方法。前者是稳定抓取所必需的，后者是对目标的期望操作所必需的。任务错误的检测需要对任务的执行方式进行在线推理。这种推理是将任务的高级象征意义与手指和被操纵对象的低级运动结合起来进行的。为了使推理成为可能，开发了用于机械装配的多指手的运动原语。钉插入、螺钉插入和螺钉旋转已成功实现为运动原语。在多指手的机械装配实验中，由于装配过程中存在不可预测的摩擦力，常常会在指尖与被操纵物体接触处发生滑移。这个失误导致了任务的错误。为了实现错误检测和恢复，开发了一种基于实时三维视觉的任务监控系统。这是作为与任务执行控制进程并行的进程来实现的。在最小的时间延迟内检测被操纵对象的姿态误差，同时又不损失精细灵巧的操纵能力。

A dexterous manipulation system has been developed as a test bed for detecting errors of task : the system is conposed of a multi-fingered hand equipped with force/torque sensors on each fingertips, a 3-D realtime stereo vision to measure the pose of a manipuiated object, and a realtime figer controller.Following functions have been developed :(1).The grasped object is manipulated through rolling contacts between fingertips and the surface of the object enabling fine and control with 10ms of servo cycle time.(2).The 3-D stereo vision measures the pose the pose of the grasped object in every 33ms.(3).A method of separating internal forces and external forces from the force/torque sensorsigna1s has been developed. The former is necessary for the stable grasp and the latter is required for desired manipulation of the object.Detection of task errors requires online reasoning of how the task is being executd. This reasoning is made by combining the higher level synbolic meaning of the task and the lower level motion of fingers and the manipulated object. To enable the reasoning, motion primitives of the multi-fingered hand for mechanical assembly have been developed. Peg-insertion, screw-insertion, and screw-rotation have been successfully implenented as motion primitives.In the mechanical assenbly experinents by multi-fingered hand, a slip is very often caused at the contact of the fingertip and the manipulated object mainly because of an unpredictab1e friction force in the assenbly. This slip leads to an error of the task. For the error detection and recovery, a task monitoring system has been developed using the real time 3-D vision. This is implemented as a parallel process with the task execution control process. Detection of errors in the pose of the manipulated object in a minimum time delay without losing fine and dexterous manipu1ation capability.