Assembly Plan From Observation

摘要

Abstract Currently, a major bottleneck to the widespread use of robots is the time consuming nature of programming and debugging a desired sequence of operations of a robot system. Further, the number of trained people who can program robots is also limited. If we can devise self-programming systems that “learn” (program themselves) by observing a human performing the same task, then we would have a situation where nonprogrammers (i.e., manufacturing engineers) can define and debug any process they choose. We have been developing an assembly plan from observation (APO) system that can learn robot programs through observation. In the APO system, a human operator performs assembly tasks in front of a video camera. The system obtains a continuous sequence of images from the camera, which records the assembly tasks. In order for the system to recognize assembly tasks from the sequence of images, the system has to perform the following six operations: