[Advances in Artificial Intelligence]

By

Advances In Artificial Intelligence

Introduction

Robot motion planning systems have used many space and object representations. Objects have been modeled by polygons and polyhedra or bounded by curved surfaces. Free space has been partitioned into Vornoi regions or, heuristically, free corridors(Elfes, 2012 , pp. 67-71). Traditionally, the models have been hard edged; positional uncertainty, if considered at all, was used in just a few special places in the algorithms, expressed as a Gaussian spread. Partly, this over simplification of uncertainty information is the result of analytic difficulty in manipulating interacting uncertainties, especially if the distributions are not Gaussian. Incomplete error modeling reduces positional accuracy. Seriously, it can produce entirely faulty conclusions: A false determination of an edge in a certain location, for instance, can derail an entire train of inference about the location or existence of an object. Because they neglect uncertainties and alternative interpretations, such programs are brittle. When they jump to the right conclusions, they do well, but a small error early in the algorithm can be amplified to produce a ridiculous action(Serey, 2011 , pp. 128-129). Most AI-based robot controllers have suffered from this weakness. The Mobile Robot Laboratory (MRL) at Carnegie-Mellon University has built their share of brittle controllers. Occasionally, however, MRL stumbled across numeric (as opposed to analytic) representations that seem to escape this fate. One numeric representation is deep inside the program that drove the Stanford Cart in 1979. Each of 36 pairings of nine images from a sliding camera produced a stereo depth measurement of a given feature, identified by a correlator, in the nine images. Some pairings were from short baselines and had large distance uncertainty; others were from widely separated viewpoints with small spread. The probability distributions from the 36 readings were combined numerically in a IOOO-cell array, each cell representing a small range interval. Correlator matching errors often produced a multi peaked resultant distribution, but the largest peak almost always gave the correct range. The procedure was the most error-tolerant step in the Cart navigator, but it alone did not protect the whole program from brittleness (Thorpe, 2010 , pp. 501-503).

A robot should be able to navigate around a space with some persistent memory of the features of that space. In my system, the robot begins by taking a sonar sounding, which produces a polar distance map of the robot's immediate neighborhood. The robot is assumed to be at the origin (0,0), and these initial soundings are taken to be the robot's initial map. Then the robot proceeds to move in some direction (goal planning is outside the scope of this project), stops, and takes another sounding. This sounding is fit to the existing map, on the assumption that the features in the robot's neighborhood have not changed much. The best fit returns a most likely location of the robot relative to the origin; the soundings are then shifted by the robot's now-known position, and contributed to the map. This cycle repeats indefinitely as the robot explores; at each stop, ...