Related Experiment Video
Updated: Jul 17, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
568
Deep reinforcement learning-aided autonomous navigation with landmark generators.
Xuanzhi Wang1, Yankang Sun1, Yuyang Xie1
1Department of Integrated Circuit Science and Engineering, Nanjing University of Posts and Telecommunications, Nanjing, China.
Frontiers in Neurorobotics
|September 7, 2023
Summary
This study introduces a novel robot navigation framework combining traditional planning with deep reinforcement learning (DRL) to overcome dynamic obstacles. The new method significantly improves navigation efficiency and safety, reducing training time and preventing robots from getting stuck.
Area of Science:
- Robotics
- Artificial Intelligence
- Machine Learning
Background:
- Mobile robots are crucial in various applications, but accurate autonomous navigation, especially with dynamic obstacles, remains a challenge.
- Traditional navigation methods struggle with dynamic environments, while deep reinforcement learning (DRL) faces issues like long training times and lack of long-term memory, hindering real-world application.
- Existing DRL approaches for robot navigation can be inefficient and prone to getting robots stuck in dead ends.
Purpose of the Study:
- To develop an improved autonomous navigation framework for mobile robots that addresses the limitations of traditional methods and DRL.
- To reduce the training time and enhance the robustness of DRL-based navigation systems.
- To enable crash-free navigation in complex environments with numerous and fast-moving dynamic obstacles.
Main Methods:
- A hybrid navigation framework integrating traditional global path planning with DRL-based local planning.
- Utilizing DRL to navigate towards high-value landmarks identified on the global path.
- Incorporating a feature extraction network with memory modules to preserve long-term dependencies in DRL.
Main Results:
- The proposed method demonstrated superior performance compared to traditional and end-to-end DRL navigation approaches.
- Achieved an average of 20% improvement in navigation efficiency (time and path length) and 34% in safety (fewer collisions).
- Showcased a 26.6% higher success rate and strong robustness in scenarios with many rapidly moving dynamic obstacles.
Conclusions:
- The hybrid navigation framework effectively combines global and local planning for efficient and safe mobile robot navigation.
- The integration of memory modules in DRL enhances its capability to handle long-term dependencies, crucial for complex navigation tasks.
- This approach significantly reduces training difficulty and improves the practical applicability of DRL in real-world robot navigation scenarios.
