Simultaneous map building and localization
At the IROS '91 workshop in Osaka (3-5 November 1991) John Leonard of the NEC Research Institute and Hugh Durrant-Whyte of Oxford named an open problem of mobile robotics, simultaneous map building and localization, and defined it as long-term globally referenced position estimation without a priori information. The problem is hard because of a paradox: to move precisely a robot needs an accurate map, and to build an accurate map it must know precisely where it sensed from. For sonar they fit the vehicle with several servo-mounted sensors, learn a subset of environment features from the initial location and then track them.
Why it matters
The problem from which the term SLAM grew is posed here for the first time as one problem rather than two: map or motion, chicken or egg. Everything after, from MonoSLAM in 2003 to the navigation of today’s robots, answers this formulation.
Proceedings IROS '91: IEEE/RSJ International Workshop on Intelligent Robots and Systems '91, pages 1442-1447 of volume 3. The record rests on the paper's abstract, because the full text is paywalled; the abstract carries no figures, and the record claims neither the size of the environment, nor a positioning accuracy, nor the number of sonars. The abbreviation SLAM does not occur in the abstract: the problem is called simultaneous map building and localization here, and the abbreviation settled later, for which this package has no date. Leonard is given as with the NEC Research Institute in Princeton, Durrant-Whyte with the Department of Engineering Science at Oxford.