Multithreaded Asynchronous Deep Reinforcement Learning With Multisensor Fusion for Robot Collision Avoidance.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Developing safe and efficient navigation for robotic vehicles in dynamic environments is crucial.
• Existing methods often struggle with complex, unpredictable scenarios.

Purpose of the Study:

• To present a novel collision-avoidance method for robotic vehicles using deep reinforcement learning (DRL).
• To improve navigation safety and efficiency in dynamic environments.

Main Methods:

• A multithreaded asynchronous proximal policy optimization (MAPPO) for efficient offline training.
• A multisensor fusion measurement (MSFM) combining global reference path (GRP), laser scanner measurement (LSM), and motion energy (ME).
• A collision-avoidance neural network (CANN) and a premature collision prediction (PCP) module for enhanced threat detection and safety.

Main Results:

• The MAPPO method significantly improves sample efficiency during policy learning.
• The MSFM and CANN effectively generate obstacle features for threat assessment.
• Extensive experiments demonstrate successful collision avoidance and farsighted navigation in complex simulated and real-world scenarios.

Conclusions:

• The proposed DRL-based collision-avoidance method is effective and robust for robotic navigation in dynamic environments.
• The system enables robots to make intelligent, farsighted decisions to prevent collisions while progressing towards their objectives.