Deep Reinforcement Learning for Autonomous Driving with an Auxiliary Actor Discriminator

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Path planning and obstacle avoidance are critical for intelligent robots, particularly in unknown dynamic environments.
• Existing methods often struggle with the flexibility and rapid decision-making required in such complex scenarios.

Purpose of the Study:

• To develop an advanced robot navigation system capable of effective path planning and obstacle avoidance.
• To enhance robot decision-making and exploration capabilities in unknown dynamic environments.

Main Methods:

• A state attention network (SAN) was developed for feature extraction of robot-obstacle interactions.
• An auxiliary actor discriminator (AAD) was implemented to calculate collision probabilities.
• Goal-directed and gap-based navigation strategies, guided by heuristic knowledge (HK), were employed.
• The Soft Actor-Critic (SAC) algorithm was used for policy training in simulated environments.

Main Results:

• The proposed approach demonstrated convergence towards optimal action strategies for robot systems.
• Robots explored unknown environments with significantly fewer moving steps, showing a decrease of 33.9%.
• The system achieved higher average rewards, with an increase of 29.15% compared to other methods.

Conclusions:

• The developed navigation system enhances robot performance in complex, unknown dynamic environments.
• The integration of SAN, AAD, and heuristic knowledge offers a robust solution for intelligent robot exploration.
• This research contributes to more efficient and effective autonomous robot navigation.