A Comparison Study between Traditional and Deep-Reinforcement-Learning-Based Algorithms for Indoor Autonomous...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Autonomous navigation for mobile robots necessitates a choice between traditional control and artificial intelligence (AI) algorithms.
• This decision is complex, influenced by robot computational capacity, sensor data, and environmental dynamics.
• Selecting appropriate algorithms is crucial for effective navigation, especially in complex, changing environments.

Purpose of the Study:

• To review and identify suitable autonomous navigation algorithms for mobile robots operating in dynamic environments.
• To compare traditional algorithms with Deep Reinforcement Learning (DRL)-based approaches.
• To provide recommendations for algorithm selection based on robot development requirements.

Main Methods:

• A comprehensive review of autonomous navigation algorithms was conducted.
• Selected algorithms, including Dynamic Window Approach (DWA), Timed Elastic Band (TEB), CADRL, and Soft Actor-Critic (SAC), were compared.
• Performance evaluation was carried out on a robotic platform to assess advantages and disadvantages.

Main Results:

• The study identified specific traditional and DRL-based algorithms best suited for dynamic environments.
• Comparative analysis revealed distinct performance characteristics, advantages, and disadvantages for DWA, TEB, CADRL, and SAC.
• Results indicate that the optimal algorithm choice is contingent upon the specific application and robot characteristics.

Conclusions:

• No single algorithm is universally superior; selection depends on application-specific needs and robot capabilities.
• Traditional algorithms like DWA and TEB offer viable solutions under certain conditions.
• DRL-based algorithms such as CADRL and SAC demonstrate significant potential for complex dynamic navigation tasks.