Abstract
Traditional Autonomous Underwater Vehicle (AUV) missions rely on pre-scripted command sequences, which often fail to capture dynamic oceanographic phenomena. This paper presents a unified framework that integrates high-level mission planning, real-time execution, and probabilistic state estimation. By embedding Hidden Markov Models (HMMs) within a constraint-based temporal planner, the proposed system enables AUVs to autonomously detect, track, and sample elusive scientific features such as underwater particulate layers.
Introduction
The ocean is a highly dynamic and uncertain environment. Conventional AUV operations are typically "scripted," where operators define a fixed sequence of commands prior to deployment. However, if a vehicle encounters a "feature of interest"—such as a nutrient-rich plume or a specific temperature layer—a scripted vehicle will simply continue its pre-programmed path without adjusting its survey resolution or stopping to collect samples.
The primary challenge in developing more adaptive AUVs lies in bridging the gap between slow, computationally expensive high-level deliberation and fast, low-level reactive control. Most existing systems struggle to integrate these two paradigms, resulting in vehicles that are either unresponsive to environmental changes or incapable of executing complex scientific objectives.
The Teleo-Reactive Executive (T-REX) Architecture
To address these challenges, the researchers developed the Teleo-Reactive Executive (T-REX), a hierarchical control architecture. T-REX partitions the robot's control logic into several functional units called Reactors, each operating over different temporal horizons:
- Mission Manager: Oversees the entire mission (e.g., a 10-hour survey) and manages high-level goals.
- Navigator: Operates over a shorter horizon (e.g., 60 seconds) to generate specific waypoints and trajectories.
- Vehicle Control Subsystem (VCS): Handles low-level execution, such as depth control and heading.
Key Technical Innovations
- Constraint-Based Temporal Planning: Utilizing the EUROPA2 library, the system treats the mission as a network of constraints (e.g., "reach point X by time Y with >20% battery") rather than a linear list of steps.
- Integrated State Estimation: Hidden Markov Models (HMMs) are embedded directly into the planning loop. This allows the planner to treat the probability of encountering a scientific feature as a first-class state variable.
- Interleaved Synchronization and Deliberation: T-REX allows the current plan to be synchronized with incoming sensor data while the planner simultaneously works on future goals, avoiding the bottlenecks of traditional "Sense-Plan-Act" cycles.
Experimental Evaluation and Results
The system was evaluated through real-world sea trials in Monterey Bay, California, using a Dorado AUV. The trials targeted two specific phenomena: Intermediate Nepheloid Layers (INL)—sheets of suspended sediment—and Estuarine Plumes.
Key Findings
- Autonomous Feature Detection: The system successfully identified INLs in the Monterey Canyon and autonomously triggered water sampling canisters without human intervention.
- Operational Efficiency: Onboard planning was highly efficient, with an average CPU load of approximately 5%, peaking at 25% during active feature processing.
- High-Level Autonomy: During a four-hour survey, the system autonomously generated and dispatched 188 commands. In a traditional setup, a human operator would have had to manually pre-program these commands based on guesswork.
- Robustness: The integrated state estimator effectively filtered out sensor noise, preventing false-positive sampling events by ensuring detections matched the expected scientific profiles.
Discussion and Future Work
A critical insight from this work is the value of decoupling plan synchronization from deliberation. This separation allows the robot to remain reactive to sensor noise without compromising its long-term objectives. Furthermore, using a unified declarative model for both high-level mission management and low-level instrument control minimizes architectural mismatches.
While successful, future work aims to address several limitations:
- Scalability: Testing the system's performance over multi-day missions.
- Online Learning: Enabling the robot to refine its HMM parameters and models of scientific features in real-time during a mission.
- Multi-Agent Coordination: Extending the T-REX framework to coordinate fleets of heterogeneous assets, such as gliders and seafloor stations, for comprehensive 4D ocean mapping.
Summary updated