This paper describes the design of a system for handling analysis of sensory input. This analysis could be used for guiding robot motion, but this application is not the main thrust of the paper. Similarly, the paper does not furnish any detail on any algorithms which are invoked to perform any of the analysis, but instead gives a high-level view of function and database organization.

The sensory system holds three-dimensional representations (oct-trees) of objects and uses them to produce two-dimensional projections of what can be expected. Recognizer modules attempt to match what is “seen” with what has been predicted. Unmatched features are compared with all the object models. The processes, therefore, are divided into “predictive processes (world modelers), processes that analyze sensor data, processes that match sensor data with the models, and processes that use the sensed data and the predictions to describe the world.”

The paper describes the current status of the implementation. The fact that the implementation is not complete means that performance cannot be discussed. The authors include detailed comparisons of this system with the ACRONYM system of Brooks [1] and a system by Crowley [2]. The reference list seems heavily weighted to works from the National Bureau of Standards (NBS), but that is to be expected.

This paper is consistent with many other papers from the Automated Manfacturing Research Facility (AMRF) of the NBS in stressing hierarchical organization. It has adequate detail on the organization though not on the algorithms, and the comparisons with other systems are quite welcome. The paper is well written and organized. The incompleteness of the implementation, however, is also typical news from AMRF. In summary, the reader is still left with a vague feeling that not much substance is behind the “system” and the system itself has not been tested.