Towards the portability of knowledge in reinforcement learning-based systems for automatic drone navigation.

Area of Science:

• Artificial Intelligence
• Robotics
• Cyber-Physical Systems

Background:

• Knowledge transfer is a key challenge in artificial intelligence (AI), limiting AI model applicability across diverse tasks and environments.
• Cyber-Physical Systems (CPS) require enhanced reliability and autonomy, often necessitating adaptable AI solutions.
• Unmanned vehicles (drones) in CPS face complex navigation challenges in dynamic, obstacle-filled environments.

Purpose of the Study:

• To develop and evaluate a reinforcement learning system for unmanned vehicles capable of knowledge transfer.
• To enable autonomous navigation of drones in unknown environments by leveraging learned knowledge from other agents.
• To quantify the effectiveness of knowledge portability in improving navigation success rates and reducing learning times.

Main Methods:

• Implementation of a reinforcement learning system equipping unmanned vehicles (drones) for autonomous navigation.
• Development of a method to isolate and transfer learned navigation knowledge between AI agents.
• Experimental validation of the knowledge transfer system in various simulated environments with physical obstacles.

Main Results:

• The proposed system demonstrated successful knowledge isolation and transfer between agents.
• Agents utilizing transferred knowledge achieved higher success rates in navigating unknown environments.
• A significant reduction in learning time was observed for agents employing transferred knowledge compared to baseline methods.
• The system outperformed the baseline in 78.3% of tests, showing improved success rates and reduced learning times.

Conclusions:

• The developed reinforcement learning system effectively enables knowledge portability for unmanned vehicles in cyber-physical systems.
• Knowledge transfer significantly enhances the autonomy and reliability of drones navigating complex, unknown environments.
• This approach offers a promising solution for improving AI adaptability and efficiency in real-world applications.