Accession Number : AD1051419


Title :   An Initial Approach for Learning Objects from Experience


Descriptive Note : Technical Report,01 Oct 2016,08 Feb 2018


Corporate Author : US Army Research Laboratory Aberdeen Proving Ground United States


Personal Author(s) : Kelley,Troy D ; McGhee,Sean M ; Milton,Jonathan


Full Text : https://apps.dtic.mil/dtic/tr/fulltext/u2/1051419.pdf


Report Date : 02 May 2018


Pagination or Media Count : 34


Abstract : The US Army Research Laboratorys Vehicle Technology Directorate (VTD) and the Human Research and Engineering Directorate, as part of VTDs 6.1 refresh program, have initiated a program called Adaptive Perception Processes for Learning from Experience (APPLE). The programs goal is to develop a set of perception capabilities that are sufficient to enable continuous object learning, where new object instances and categories can be learned from experience in an open-set framework. We have shown preliminary results from initial tests using a motion detection algorithm to delineate objects which are then fed to a simple feed-forward neural network without any other processes in the pipeline. Our neural network trains to an asymptote of acceptable performance and our topology can be defined using our nonparametric model. ARL will continue research to determine the best algorithms to use in the pipeline for the APPLE program.


Descriptors :   neural nets , perception , learning , robotics , motion detectors


Subject Categories : Cybernetics


Distribution Statement : APPROVED FOR PUBLIC RELEASE