End-to-End Autonomous Navigation Based on Deep Reinforcement Learning with a Survival Penalty Function.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous navigation in dynamic, map-less environments presents significant challenges.
• Traditional methods struggle with sparse rewards and complex obstacle avoidance.

Purpose of the Study:

• To propose an end-to-end deep reinforcement learning (DRL) approach for autonomous navigation.
• To enable nonholonomic wheeled mobile robots (WMRs) to navigate dynamic, map-less environments effectively.
• To address the sparse reward problem and ensure collision-free path planning.

Main Methods:

• Utilized two actor-critic (AC) frameworks: deep deterministic policy gradient (DDPG) and twin-delayed DDPG (TD3).
• Introduced a comprehensive reward function incorporating a survival penalty to guide the WMR towards its target.
• Connected consecutive episodes to increase cumulative penalties for obstacle scenarios, preventing training failure.

Main Results:

• Simulations in various scenarios (obstacle-free, parking lot, intersections, multiple obstacles) demonstrated the method's efficiency and safety.
• The TD3 algorithm showed faster convergence and greater stability during training compared to DDPG.
• TD3 achieved a higher task execution success rate during evaluation.

Conclusions:

• The proposed DRL approach with a survival penalty function effectively enables autonomous navigation in challenging environments.
• The TD3 algorithm offers superior performance in terms of training efficiency and navigation success rate over DDPG.
• This method provides a robust solution for collision-free path planning for WMRs.