CBMA-ViNet: A Path Planning Network Integrated with Attention Mechanism and Multi-Level Abstract Maps

This paper explores an improved path planning method based on Value Iteration Networks (VIN): CBMA-ViNet. This method combines attention mechanisms and multi-level abstract map mechanisms to optimize the reward function module in traditional VIN. Traditional Value Iteration Networks generate a reward map through a feature extraction network and perform value iteration on this map to learn policies. This process enables the network to flexibly adapt to different environments, learn complex decision rules, and effectively handle obstacle avoidance problems. However, VIN exhibits unstable performance and difficulty in utilizing long-range information in large maps and complex local environments. To address these issues, this study introduces the Convolutional Block Attention Module (CBAM), a lightweight and general-purpose attention module that enhances the network's expressive power. Additionally, this study proposes the Multi-Level Abstract Value Iteration Network (MutiA-VIN), which describes states at different resolutions and includes both global and local information, thereby effectively improving performance in large planning spaces. The CBMA-ViNet method combines CBAM and MutiA-VIN, enhancing the stability and accuracy of the reward function in large maps while strengthening the network's adaptability to complex environments and the algorithm's speed. Through these innovative improvements, CBMA-ViNet has become a powerful path-planning tool suitable for applications such as autonomous driving and robotic navigation.