Advancing Self-Localization and Intelligent Mapping (SLIM) for Swarm of Autonomous Unmanned Underwater Vehicles using machine learning
Abstract
Abstract The mission of the Naval Sea Systems Command (NAVSEA) is to “Keep our Navy #1 in the world by providing superior, cost e.ective, and innovative logistics, engineering, information technology, and quality assurance solutions that meet the life-cycle requirements of our Navy.” [1]. NUWC Division Keyport focuses on delivering innovative solutions for today’s warfighter. While there are likely a multitude of defense strategies being deployed to achieve such a mission, one consistent theme is the swarm of multiple sensory units for improving perception, localization, and autonomous planning and decision making. In most “multi-agent system” problems, the key idea is to extract the meaningful information from individual agents, and fusing/integrating such information to obtain a complete world model. In modern Unmanned Underwater Vehicles (UUVs), data is collected from a multitude of sources within the swarm. In swarm of UUVs, the agents each have local world views and observe only fraction of the world model. The overall task is then to fuse/integrate their individual data to construct an integrated, global view of their operational space. While world view modeling through swarm data integration has seen significant advancements within the “ground-based” and “aerial-based” robotic systems, the field within UUV systems trails behind due to the fact that significant challenges arise when moving underwater. In particular, distributed state estimation in underwater environments su.ers due to either lack of communication between agents, and/or unreliable networks resulting in packets being dropped, communication delays, and total link failure/timeouts. Moreover, integrating individual agent data (that may be incomplete during transfer) from multiple autonomous/semiautonomous agents is difficult in underwater environments because both the local and global measurement are extremely susceptible to noise. As a result, the current proposal has the following research objective: Research Objective Expand the Navy’s UUV capabilities through artificial intelligence to static undersea sensors and/or dynamic undersea groups to improve autonomous perception and data fusion in an e.ort to generate world models from individual sensing while localizing the sensors and UUV swarm within the model. To meet this objective, we propose to extend our prior work in pose estimation [2–12] and collaborative localization and mapping (CLAM) [13, 14] develop for unmanned ground vehicle mapping in ill-posed environments to self-localization and intelligent mapping (SLIM) in underwater environments. Toward this end we aim to investigate three fundamental advancements in deep learning and multi-agent swarms: 1) investigate Flow NET (a deep convolutional neural network (CNN) used for optical flow evaluation) and Depth NET (a deep CNN used for point-cloud estimation from multiple measurements) for fusion/integration of SONAR images from individual agents; 2) investigate novel methods for distributed state estimation using recent advances in algebraic graph theory; and 3) investigate reinforcement learning techniques using experience replay-adaptive dynamic programming (ER-ADP) for integrated swarm navigation and world model planning. Meeting this objective will provide the U.S. Navy with advanced algorithms for generation of world models using underwater perception and self-localization useful in autonomous planning and decision making.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 13, 2025
- Source ID
- N001742010007
Entities
People
- Hedi Fekmandi
Organizations
- South Dakota School of Mines and Technology
- United States Navy