A Scalable Architecture for Multi Agent Vision Based Robot Scavenging

摘要

This paper outlines the technical details and presents early results from our multi agent robot scavenging sys-tem, as demonstrated at the 2006 AAAI Robot Exhibi-tion and Scavenger Hunt Competition. Our system is designed to be easily scalable in the number of agents, which are currently limited to Evolution ER1 laptop robots capable of monocular vision. Each individual robot possesses the ability to localize itself, recognize a set of objects, and communicate with peer robots to coordinate exploration. Overview The robot scavenger problem, as posed in the 2006 AAAI Robotics Competition, is to search the conference environ-ment for a list of objects and report their position according to some map of the environment. For convenience, teams were provided with floor plans for the competition spaces. However, since the conference environment is highly dy-namic, with people constantly moving about, as well as ex-hibits, posters, and other typically characteristic environ-mental items often shifting, agents acting in the space re-quire robust reasoning algorithms in addition to simply uti-lizing floor plans. In addition, conference spaces are large and typically require long periods to explore for slow mov-ing robots. In an attempt to minimize exploration time, we have cho-sen a scalable multi agent approach to the scavenging prob-lem, that focuses strongly on the spatial reasoning, planning, and object recognition elements of the task. We are also par-ticularly interested in exploring low cost robotics and the potential of vision as a primary sensing modality, which has led to our choice of platform. We believe that vision could prove to be a flexible sensor, able to be more robust in a highly dynamic environment. Platform Our chosen platform is the ER1 robot from Evolution Robotics (Fig. 1). It serves essentially as a laptop periph-eral, connecting all sensors and motors through USB. The