COLREGS-Compliant Autonomous Collision Avoidance Using Multi-Objective Optimization with Interval Programming

Abstract

High contact density environments are becoming ubiquitous in autonomous marine vehicle (AMV) operations. Safely managing these environments and their mission greatly taxes platforms. AMV collisions will likely increase as contact density increases. In situations where AMVs are not performing a collaborative mission but are using shared physical space such as multiple vehicles in the same harbor, a high demand exists for safe and e cient operation to minimize mission track deviations while preserving the safety and integrity of mission platforms. With no existing protocol for collision avoidance of AMVs, much e↵ort to date has focused on individual ad hoc collision avoidance approaches that are self-serving, lack the uniformity of fleet-distributed protocols, and disregard the overall fleet e ciency when scaled to being in a contact-dense environment. This research shows that by applying interval programming and a collision avoidance protocol such as the International Regulations for Prevention of Collisions at Sea (COLREGS) to a fleet of AMVs operating in the same geographic area, the fleet achieves nearly identical e ciency concurrent with significant reductions in the collisions observed. A basic collision avoidance protocol was analyzed against a COLREGS-based algorithm while parameters key to collision avoidance were studied using Monte Carlo methods and regression analysis of both real-world and simulated statistical data. A testing metric was proposed for declaring AMVs as "COLREGS-compliant" for at-sea operations. This work tested five AMVs simultaneously with COLREGS collision avoidance-the largest test known to date.

A planned approach to high collision risk area

“…Unless new rules are written and adopted for USVs, existing algorithms must be compliant to established protocol for collision avoidance (i.e., COLREGS). Furthermore, the COLREGS protocols these algorithms seek to imitate have clear cut guidance only for single ship-to-ship encounters in canonical geometries, but most real world collisions arise in situations where more than two ships are involved [49]. Therefore, existing algorithms do not yet produce results comparable to experienced human operator in maneuvers that are safe and predictable [50,30].…”

“…In COLREGS, there are three canonical geometry scenarios for single ship-to-ship encounters: head-on, over-taking, and crossing. For canonical geometries, mariners and autonomous systems can arrive at a straight forward heading/speed solution in partial adherence to COLREGS [49,25,31,35,39]. Vexing situations arise when more than two ships encounter each other, and the pair-wise geometry between those ships dictate conflicting maneuvers.…”

“…The overarching algorithm of the planned phase seeks to emulate a mariner who changes after a preset threshold is reached [11,23,31,29,38,49]. Moderate realists and moderate constructionist models that value expert judgments import mariners' maneuvering and risk assessment to some extent, but suffer from the scope of expert data collection [20,36,28,40].…”

“…In fact, the AvdColregs behavior makes no claim to be fully COLREGS compliant, and only seeks to pursue maneuvers that observe COLREGS rules 13-17, a set of rules that governs one-on-one vessel interaction in canonical geometry, i.e. head-on, crossing, and overtaking [8,49]. Woerner [49]…”

“…An efficient global path planner considers static layouts such as land, buoys, and navigational hazards to form a suitable path often represented by waypoints constrained by mission requirements [32]. A local path planner, synonymous with collision avoid-ance algorithms in the maritime context, maneuvers the USV to safety in reaction to moving obstacles such as another vessel, or a newly detected hazard, often having to temporarily disregard the requirements of the mission [49]. While both essential, the studies of global path planning and collision avoidance have diverged.…”