Abstract
This paper explores the adaptation of a situated reasoning architecture, originally developed for terrestrial robots, to the domain of autonomous underwater vehicles (AUVs). Utilizing the GAPPS/REX programming language, the researchers implemented a synchronous, layered control system that integrates high-level mission directives with low-level, physics-based control algorithms. Experimental results in a controlled test tank demonstrate robust obstacle avoidance, precise station-keeping, and successful autonomous wandering, proving the portability of reactive architectures to complex 3D underwater environments.
Introduction
Operating autonomous vehicles underwater presents unique challenges, including unpredictable currents, poor visibility, and the high inertia of the vehicle itself. Traditional AI planning often struggles with the real-time demands of such environments, while pure control-theoretic approaches lack the high-level reasoning required for complex missions. This work addresses the architectural integration of these two paradigms, testing whether a reactive architecture can provide the necessary "intelligence" and "speed" for underwater exploration.
Architectural Design and Methodology
The core of the system is a situated reasoning architecture implemented in GAPPS/REX, a language designed for "situated automata." This approach compiles high-level specifications into synchronous digital circuits, ensuring a constant and predictable cycle time—a critical requirement for stable vehicle control.
Layered Competences and Hybrid Control
The architecture is structured into layered competences, inspired by subsumption theory. Lower layers focus on immediate survival tasks, such as obstacle avoidance, while higher layers manage mission-level goals like waypoint navigation. A unique "valve" mediation mechanism allows these layers to merge directives; for instance, the obstacle avoidance layer can subtly adjust a mission-driven heading to steer clear of a wall without aborting the overall goal.
To handle the complex dynamics of the underwater environment, the researchers employed a hybrid approach. Physics-heavy algorithms, such as adaptive sliding control for vehicle dynamics, were "outboarded" as low-level competences. The reactive system interacts with these algorithms as "smart actuators," providing them with goal points while the algorithms handle the low-level thruster management.
Sensing and Perception
To provide the vehicle with a 360-degree "safety bubble," a custom Proximity Obstacle Detection System (PODS) consisting of 14 sonars was integrated. This sensor array feeds into the perception layers, which build maps of proximities and space occupancy to inform the control layers.
Experimental Setup and Evaluation
Evaluation was conducted at the Deep Submergence Laboratory (DSL) at Woods Hole Oceanographic Institute (WHOI). The primary testbed was a specialized physical tank equipped with a transponder-based navigation system (SHARPS) and real-time sensor feeds.
The system's performance was tested in two primary modes:
- Pilot-Aided Mode: A human operator provided high-level goals via a joystick, while the robot autonomously managed station-keeping and obstacle avoidance.
- Autonomous Wandering: The robot navigated the tank independently, avoiding walls and experimental obstacles like pylons and ladders.
Key Insights and Results
The experiments confirmed that reactive logic is highly portable across domains. The "wander" behavior developed for ground robots translated effectively to the 3D underwater environment. Key results include:
- Precision and Reliability: The system maintained station-keeping within a remarkable 0.015 meters and consistently kept a safety buffer of 0.6 meters from obstacles.
- Safety through Synchronicity: By running the software at a fixed frequency of 3 Hz, the researchers could mathematically guarantee that the robot would have sufficient time to react to obstacles at its maximum speed of 0.7 m/s.
- Robustness: The "valve" mediation proved effective in allowing the robot to follow human or mission directives while prioritizing safety, preventing collisions even when a pilot attempted to steer into a wall.
Limitations and Future Directions
While successful, the current system relies on relatively simple reaction plans and low-resolution sonar data. Future work aims to integrate more sophisticated sensors, such as high-resolution scanning sonars or cameras, to enable more complex missions like large-scale ocean mapping. Additionally, the challenges of communication latency in deep-sea operations remain a significant area for further research.
Summary updated