Dynamic robotic tracking of underwater targets using reinforcement learning

Area of Science:

• Marine robotics
• Deep-sea ecology
• Artificial intelligence in oceanography

Background:

• Autonomous underwater robots offer scalable ocean observation but require advanced techniques for complex tasks.
• Tracking mobile underwater targets is crucial for studying marine animal connectivity and deep-sea population distribution, representing a significant knowledge gap.
• Current methods for sensor placement and path planning in dynamic underwater environments are challenging.

Purpose of the Study:

• To investigate the application of reinforcement learning for optimizing range-only underwater target tracking.
• To evaluate the effectiveness of reinforcement learning as a path planning system for autonomous surface vehicles tracking mobile underwater targets.

Main Methods:

• Implementation of a reinforcement learning method as a path planning system for an autonomous surface vehicle.
• Evaluation using an open-source model, performance metrics in simulated environments, and over 15 hours of at-sea field experiments.

Main Results:

• Demonstrated the successful application of deep reinforcement learning for autonomous underwater target tracking.
• Validated the approach through simulations and extensive real-world ocean experiments.

Conclusions:

• Deep reinforcement learning is a powerful tool for enhancing the capabilities of autonomous robots in marine environments.
• This approach encourages the deployment of advanced algorithms for effective monitoring of marine biological systems.