Hybrid Artificial Potential Fields and Spatio-Temporal Transformers for Real-Time AUV Path Planning

Technical Summary: Hybrid Artificial Potential Fields and Spatio-Temporal Transformers for Real-Time AUV Path Planning

Problem Statement
Autonomous Underwater Vehicles (AUVs) operate in complex, unstructured, and often GPS-denied environments characterized by unpredictable hydrodynamic currents, limited visibility, and stringent energy constraints. While efficient path planning is critical for mission success and safety, existing algorithms face significant trade-offs. Classical graph-search methods (e.g., A*, Dijkstra) guarantee collision-free paths but suffer from high computational latency and suboptimal path lengths in high-dimensional spaces. Sampling-based approaches (e.g., RRT*) improve exploration efficiency but often produce jagged trajectories requiring post-processing. Metaheuristic algorithms (e.g., PSO, GA, ACO) offer global search capabilities but introduce excessive trajectory complexity and are sensitive to parameter tuning, making them less suitable for strict energy budgets. Conversely, learning-based architectures demonstrate adaptability but often lack interpretability, require extensive training data, and necessitate safety fallback mechanisms when deployed in unseen scenarios. The core challenge is developing a path planning architecture that simultaneously satisfies global optimality, real-time responsiveness, kinematic feasibility, and robustness to environmental uncertainty without overwhelming limited onboard computational resources.

Methodology
The paper proposes a comprehensive comparative evaluation of thirteen state-of-the-art path planning algorithms across five distinct navigation scenarios (Diagonal, Cross, Medium, Reverse, and Long/Long Diagonal) using high-resolution underwater terrain maps (992 × 992 grid). The central contribution is a novel Hybrid APF + Spatio-Temporal (ST) Transformer framework designed to bridge the gap between reactive local control and global predictive planning.

The methodology integrates three primary components:

1. Artificial Potential Field (APF): Serves as the primary control mechanism, modeling the agent as a particle subject to virtual forces. An attractive force pulls the agent toward the goal, while a repulsive force, derived from the terrain gradient, pushes the agent away from obstacles. The motion is updated using a momentum-based formulation to ensure smooth convergence.
2. Spatio-Temporal Transformer: This component captures trajectory dynamics over time. It receives the partial trajectory and a context vector (start and goal positions), embedding them into a higher-dimensional feature space. Using self-attention mechanisms, the transformer predicts a refined next waypoint, providing a learned guidance signal that complements the APF.
3. Hybrid Control Strategy: The final motion direction is a weighted combination of the APF force, the transformer-predicted direction, and a direct goal-oriented direction. The weighting coefficients prioritize safety (APF) while benefiting from learned trajectory refinement.

To address long-distance navigation, the framework includes an adaptive attraction mechanism that scales the attractive force when the distance to the goal exceeds a threshold, and a stagnation detection system that applies random perturbations to escape local minima.

Key Contributions

• Comparative Evaluation: The study provides a rigorous benchmark of thirteen algorithms, spanning classical, sampling-based, metaheuristic, and learning-based paradigms, evaluated on path optimality, collision frequency, computational latency, and smoothness.
• Hybrid Architecture: The introduction of the Hybrid APF + ST-Transformer framework, which fuses the computational efficiency and interpretability of physics-based modeling with the adaptability of data-driven learning.
• Empirical Validation: The evaluation demonstrates that while all algorithms achieved a 100% task completion rate, the hybrid approach offers a superior balance of performance metrics compared to standalone methods.

Results
Experimental results across the five scenarios yielded the following key findings:

• Path Optimality: The Hybrid APF + ST-Transformer achieved the shortest average path length (943.15 units), outperforming standalone APF (961.70), untrained ST-Transformer (965.90), and classical methods like A* (1149.83) and Dijkstra (1190.90).
• Safety and Collision Avoidance: The hybrid method maintained a low average collision rate (0.031), significantly lower than standalone Theta* (0.378) and competitive with standalone APF (0.042). While classical methods (A*, Dijkstra, RRT*) achieved zero collisions, they did so at the cost of path length and time.
• Computational Efficiency: The hybrid approach required an average computation time of 0.96 seconds. While faster methods existed (e.g., untrained ST-Transformer at 0.112s, APF at 0.124s), the hybrid time remains within real-time constraints for most AUV missions. In contrast, metaheuristic and classical methods often exceeded 7 seconds, with some (ACO, Dijkstra) exceeding 11 seconds.
• Smoothness: Metaheuristic approaches exhibited high smoothness values (indicating oscillatory trajectories), whereas the hybrid and APF methods maintained low values, reflecting stable and continuous motion.

Significance and Claims
The paper claims that the Hybrid APF + ST-Transformer framework represents a highly viable solution for real-time AUV navigation. By harmonizing reactive obstacle avoidance with global path optimality, it addresses the persistent gap between simulation-based validation and practical deployment in resource-constrained underwater systems. The authors assert that this hybrid approach effectively reconciles the need for rapid reaction to dynamic obstacles with the generation of globally optimal trajectories, making it a principal recommendation for AUV path planning.

The study concludes with a tiered strategy for practical application:

1. Real-time Navigation: Use the Hybrid APF + ST-Transformer (or APF + A*) where a balance of speed, safety, and path quality is required.
2. Rapid Exploration/Replanning: Utilize RRT* for unknown or partially mapped environments.
3. Offline Planning: Reserve classical methods (A*, Dijkstra) for scenarios where optimality and guaranteed safety outweigh latency constraints.

The authors acknowledge limitations, noting that the evaluation was conducted on a 2D grid abstraction that does not fully capture 3D dynamics, ocean currents, or sensor noise. They suggest future work should extend these comparisons to high-fidelity simulators and hardware-in-the-loop testing.