Future Army robots will be expected to operate in highly dynamic, uncertain environments, making use of incoming data to make intelligent and independent decisions. While many autonomy applications can leverage known structure in the environment and a finite (closed) set of situations that may be encountered, the singularly uncertain domain of a battlefield demands that autonomous agents are as flexible and adaptive as possible. Within the larger scope of general adaptive systems, we seek to identify capabilities that will lead to continuous object learning (COL), to continuously learn to identify object instances and entire object categories that have never been seen before (i.e., perform open-set learning). In this report, we identify a set of core component capabilities we hypothesize to be sufficient for COL. The goal of the report is to provide a specification for the interfaces to the individual capabilities that is specific enough to derive an implementation of such an architecture, yet general enough to allow flexibility in the implementation itself. We see this specification definition as a way to ensure that well-defined, testable, and repeatable experiments can be performed when developing future autonomous systems.
展开▼