Knowledge management for autonomous systems in an open world

This paper addresses the challenge of deploying autonomous systems, focusing on their ability to make sense of, reason, and operate in the open world. Central to this endeavor is the formal representation, maintenance, and use of knowledge. The paper outlines the (i) requirements, (ii) architecture, (iii) processes, (iv) usage of knowledge bases, exemplified through an urban search-and-rescue use-case and related work from underwater military operations. First, we distill requirements from high-level desiderata (safety, efficiency, ELSA-compliance). Second, we propose a new modular knowledge base architecture, informed by existing paradigms, spanning two axes: the level of genericity (commonsense, domain, and situational knowledge) and functional scope (external, internal, and executing knowledge). Third, given the dynamic nature of open-world operations, we identify and exemplify processes for maintaining a robust knowledge base: Updating, Revising, Discarding, Merging, Matching, Translating, Inferring, Inquiring, and Learning. Fourth, we describe the overarching architecture for autonomous robotic (systems of) systems based on the OODA loop, with knowledge base at its center. In the Observe phase, data is collected via sensors; in Orient, sensor data is processed to obtain (external) situational knowledge and update (internal) performance models; in Decide, plans are made on how to perform a mission, combining executing knowledge, including operational rules, internal knowledge about skills, and external knowledge about the environment; in Act, plans are executed using available actors. This framework can guide the integration of knowledge bases in open-world systems so that future applications can extend beyond the use-case and be implemented in embodied settings with physical robots.