Deductive Reinforcement Learning for Visual Autonomous Urban Driving Navigation.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Deep reinforcement learning (RL) has seen limited application in complex, real-world scenarios like autonomous driving.
• Vision-based autonomous urban driving navigation (VB-AUDN) demands robust agents capable of handling unpredictable environments.

Purpose of the Study:

• To develop a novel deep reinforcement learning framework, Deductive Reinforcement Learning (DeRL), for enhanced safety and reliability in VB-AUDN.
• To improve the decision-making capabilities of autonomous driving agents by enabling them to anticipate future events.

Main Methods:

• Introduced a Deduction Reasoner (DR) to predict future environmental transitions using a parameterized model.
• Implemented a self-assessment mechanism within DR to evaluate policy consequences along predicted trajectories.
• Developed a Semantic Encoder Module (SEM) for robust feature extraction from raw visual input, ensuring environmental adaptability.

Main Results:

• DeRL demonstrated superior performance compared to state-of-the-art model-free RL approaches on the CARLA benchmark.
• The proposed method achieved higher success rates in goal-directed navigation tasks.
• DeRL significantly improved driving safety metrics in simulated urban environments.

Conclusions:

• DeRL offers a promising approach for advancing safe and reliable autonomous urban driving.
• The integration of predictive reasoning and robust perception enhances agent decision-making in complex driving scenarios.
• This framework provides a foundation for more sophisticated VB-AUDN systems.