End-to-End One-Shot Path-Planning Algorithm for an Autonomous Vehicle Based on a Convolutional Neural Network...

Area of Science:

• Robotics and Automation
• Artificial Intelligence
• Computer Vision

Background:

• Path planning is crucial for robot navigation and automated driving.
• Existing iterative path-planning methods can be slow, especially for large or complex environments.
• There is a need for faster and more efficient path-planning solutions.

Purpose of the Study:

• To develop an end-to-end path-planning algorithm using a fully convolutional neural network (FCNN).
• To enable efficient path planning on grid maps of varying sizes and shapes (10x10 to 80x80).
• To provide a model capable of finding both the lowest-cost and shortest paths.

Main Methods:

• Implemented an end-to-end path-planning algorithm based on a fully convolutional neural network (FCNN).
• Trained a general path-planning model considering traversability cost for grid maps.
• The FCNN model generates probability maps for lowest-cost and shortest paths, reconstructing the optimal path via highest probability selection.

Main Results:

• The proposed FCNN method demonstrates superior speed advantages over traditional algorithms.
• Achieved average optimal rates of 72.7% for lowest-cost paths and 78.2% for shortest paths.
• Reported average success rates of 95.1% and 92.5%, with average length rates of 1.04 and 1.03, respectively.

Conclusions:

• The end-to-end FCNN path-planning algorithm offers significant speed improvements for robotics and automated driving.
• The method effectively plans optimal paths, considering both cost and shortest distance.
• This approach provides a robust and efficient solution for path planning in diverse grid map environments.