In context
In 2024, as autonomous mobile robots expanded into complex outdoor environments, path planning in 3D terrain remained a bottleneck. Traditional A* algorithms suffered from inefficient heuristic modeling, especially in mountainous scenarios, prompting research into deep learning-based approaches to improve search efficiency and path quality.
What was reported
A study from Guilin University of Electronic Technology presented a deep heuristic-improved 3D A* algorithm using a dynamic field-of-view model. The method constructs a heuristic function via a convolutional neural network, enhanced by a local dynamic field-of-view that focuses on key terrain between start and goal, reducing noise from irrelevant map areas. A distance-weighted loss function improves cost estimation for long paths.
Simulations in a 100×100 grid DEM environment showed that compared to existing deep learning-based A* variants, the proposed algorithm improved path cost prediction accuracy by 45.2% on average, search efficiency by 12.8%, and path quality by 1.2%. It also outperformed empirical modeling-based A* in search speed.
The network uses a three-channel input (local map, robot position, goal) and a CNN with two convolutional layers and three fully connected layers, trained with supervised data from traditional A*.
Why it mattered
This work addressed the generalization and accuracy challenges of learning-based heuristics in 3D planning, offering a practical path toward efficient autonomous navigation for mountain robots and other outdoor industrial applications.
"The proposed algorithm can improve the prediction accuracy of path cost by 45.2% on average, the average search efficiency by 12.8%, and the average path quality by 1.2% in 3D scenarios."
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-09-15 · “基于动态视场的深度启发改进3维A*算法”