Intelligent Sensor Fusion for the 1997 AAAI Mobile Robot Competition

Abstract

The Colorado School of Mines (CSM) will field a team of undergraduates for two events in the 1997 AAAI Mobile Robot Competition: Life on Mars and HOTS D’oeuvres. The objectives of the team are (1) to gain experience with implementing behaviors under a hybrid deliberative/reactive style of architecture and (2) to transfer research being conducted at CSM in intelligent sensor fusion to new applications. The students are preparing the entries as part of the class projects for the MACS415: Introduction to AI Robotics and Computer Vision and MACS370: Field Session courses. The team is using two different robots, both running subsets of the Sensor Fusion Effects (SFX) architecture implemented in C++. The term intelligent sensor fusion denotes a broad characterization of sensor fusion, dealing with how observations from one or more logical sensors can be combined reliably into a coherent percept. Two components of our research in intelligent sensor fusion are being applied to the competition. The first component addresses how to combine the observations from multiple sensors into a percept and generate a measure of belief. This is a key element of the sensing strategy for the Hors D’oeuvres event, where evidence from simple features (audio, motion, and flesh colored regions) are combined to produce a measure of belief in the presence of a person. Our approach is an adaptation of (Goodridge 1996) for a Dempster-Shafer evidential scheme. The second component deals with combining sensor observations over time. This is an important aspect of our approach to the Find Life on Mars event, since a suspected Martian may disappear due to a temporary occlusion, because it has attempted to hide, or the sensor reported a faulty observation. In order to successfully balance resources (energy, time) with the goals (find Martians), the robot must persist in its belief that it saw a Martian for some period of time in order to continue searching, but it must eventually give up if it cannot reacquire the target. The problem is exacerbated by the lack of explicit a priori models of the Martian to guide