Maritime areas such as harbors, channels, and straits require precise piloting of vehicles as they are typically congested with other unmanned surface vehicles (USV) and hazards (i.e. buoys, bridges). Current harbor piloting systems use a human to integrate numerous inputs including his or her sense of sight and hearing. To overcome the need for a human expert, a straightforward solution for autonomous navigation would be to set thresholds that are based on the input sensor data received to statically execute a predefined set of commands for each particular situation. However, while expert systems policies are characterized by rapid initial capability, their deterministic behavior provides vulnerabilities in specific and complex scenarios. EpiSci proposes developing, demonstrating, and commercializing AlphaUSV, a hybrid yet synergistic fusion of expert systems and deep learning algorithms for safe and precise navigation of UUVs in congested maritime areas. AlphaUSV, with its expert system and deep learning integration as its core brain, provides a highly effective, reliable, and explainable system, allowing for assured and bounded behavior and reliable command and control, even with limited, degraded, or intermittent communications. EpiSci is highly qualified to develop and demonstrate the proposed AlphaUSV system with extremely low development risks as evidenced by prior and ongoing successes on similar autonomous sensing and navigation technologies such as small unmanned aerial systems (1st Prize: NSWC Crane AISUM CRANE Challenge), unmanned aerial systems (Top 3 in the DARPA ACE AlphaDogfight program), and UMAA-compliant autonomous UUV technologies (AlphaUUV feasibility demonstration via digital twin). Consequently, we plan to leverage in-house, state-of-the-art computer vision and deep reinforcement learning (DRL) algorithms to detect objects using sensors such as cameras and lidar, estimate with precision their trajectories, and plan for a safe path with high confidence. The problem described in the AlphaUSV program is analogous to the AISUM CRANE Challenge but in an maritime domain.? We will leverage our winning solution in the CRANE challenge for this program. Furthermore, EpiSci can leverage real-world experience from a UCSD collaboration with Naval Information Warfare Center (NIWC) Pacific through the Naval Research Enterprise Internship Program (NREIP) focusing the development on key aspects for the design of autonomous navigation for USV.
Benefit: To overcome the need for a human expert, a straightforward solution for autonomous navigation would be to set thresholds that are based on the input sensor data received to statically execute a predefined set of commands for each particular situation. For example, a set of expert system (ES) rules could use the estimated position and trajectory of an oncoming ship to chart a path around it. However, while ES policies are characterized by rapid initial capability through the integration of expert knowledge, their deterministic behavior provides vulnerabilities in specific and complex scenarios. On the other hand, artificial intelligence and deep learning methods offer comprehensive capabilities in a variety of situations without dedicated human input but exhibit fragility when faced with novel situations not covered in the training environment. The technical problem of a dynamic USV mission planning tool is primarily its lack of integration of human expertise (logical reasoning in the form of ES and ontology) and deep learning. These factors present an unprecedented opportunity to advance the optimization and autonomy of USV technology to exploit the best of both expert systems and machine learning. AlphaUSV is a hybrid yet synergistic fusion of expert systems and deep learning algorithms for safe and autonomous navigation planning of unmanned surface vehicles in congested maritime areas. The proposed AlphaUSV technology innovation and tools development build on EpiScis Tactical AI paradigm has been successfully applied for several high-profile autonomous system development projects funded by various Department of Defense (DoD) agencies: DARPA (Air Combat Evolution/AlphaDogfight Trial programs), the Office of Naval Research (Tactical AI-based Autonomous Network Management), Air Force Research Laboratories (SwarmSense.ai), Armys DEVCOM (Counter-Swarm program) and Navys Naval Sea Systems Command SBIR project (AlphaUUV), to name a few. AlphaUSV, with its ES and deep learning integration as its core brain for safe and precise navigation in congested maritime areas, is aimed at providing a highly effective, reliable and explainable system. This allows assured and bounded behavior and reliable command and control, even with limited, degraded or intermittent communications. The primary benefit of adopting the modular, hybrid Tactical AI over either pure ES or deep learning models is that a naive, pure deep learning-based algorithm will likely fail to establish the necessary trustworthy relationship with human operators due to the black-box nature of deep learning algorithms. In addition, the modular system architecture of Tactical AI naturally allows to use existing navigation control mechanisms with very few changes. Such benefits drastically increase the acceptance of the resulting AI-based autonomy for UxVs.
Keywords: Maritime, Maritime, USV, Unmanned Surface Vehicles, Unmanned Aerial Systems, Tactical AI, Unmanned Underwater Vehicles, Machine Learning, UUV
