Improved Robot Path Planning Method Based on Deep Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robotic path planning is a complex nonlinear problem.
• Deep Reinforcement Learning (DRL), specifically Deep Q-Network (DQN), has shown promise but faces challenges like dimensionality, convergence, and sparse rewards.
• Existing methods struggle with efficient training and optimal path generation.

Purpose of the Study:

• To propose an enhanced Double Deep Q-Network (DDQN) path planning approach.
• To address the limitations of existing DRL algorithms in robotics.
• To accelerate model convergence, enhance training stability, and generate superior paths.

Main Methods:

• Implemented a dimensionality reduction technique on training data.
• Introduced a dual-branch network architecture for separate navigation and obstacle avoidance.
• Integrated an expert experience module for accelerated early-stage training within the Epsilon-Greedy algorithm.
• Optimized the reward function for more immediate environmental feedback.
• Utilized a two-branch network incorporating expert knowledge and optimized rewards.

Main Results:

• The enhanced DDQN algorithm demonstrated accelerated model convergence.
• Improved training stability was observed in both simulated and real-world experiments.
• The approach successfully generated smooth, shorter, and collision-free paths.

Conclusions:

• The proposed enhanced DDQN path planning method effectively overcomes limitations of traditional DRL.
• This approach offers a more robust and efficient solution for robotic navigation and obstacle avoidance.
• The algorithm shows significant potential for practical applications in robotics.