RL-DOVS: Reinforcement Learning for Autonomous Robot Navigation in Dynamic Environments.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous navigation in dynamic, populated environments is crucial for service robots.
• Reinforcement learning (RL) has shown promise in applications like autonomous driving and pedestrian navigation.

Purpose of the Study:

• To present a novel reinforcement learning dynamic object velocity space (RL-DOVS) planner for robot navigation in dynamic environments.
• To explicitly incorporate robot kinodynamic constraints into action selection.
• To improve the efficiency of the learning process for navigation.

Main Methods:

• The RL-DOVS planner utilizes an environment model representing dynamism in the robocentric velocity space as input.
• Two approaches are proposed: RL-DOVS-A (automatic action learning) and RL-DOVS-D (human driver action selection).
• The method explicitly considers robot kinodynamic constraints for action selection in each control period.

Main Results:

• The use of dynamic obstacle velocity information significantly speeds up the training process compared to methods using raw sensor data or basic obstacle information.
• Evaluations in diverse scenarios with varying numbers of static and dynamic agents demonstrate the planner's effectiveness.
• Performance is benchmarked against state-of-the-art navigation techniques.

Conclusions:

• The RL-DOVS planner offers an efficient and effective solution for autonomous robot navigation in complex, dynamic environments.
• Representing environmental dynamism in the robocentric velocity space is key to accelerated learning.
• The proposed method advances the capabilities of service robots in real-world populated settings.