One-Shot Multi-Path Planning Using Fully Convolutional Networks in a Comparison to Other Algorithms.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Path planning is essential for robot arm movement and navigation.
• Existing methods are iterative and struggle with multi-agent systems.
• Centralized approaches are computationally intensive and do not scale well.

Purpose of the Study:

• To introduce a novel, one-shot path planning method using convolutional neural networks.
• To enable efficient path generation for single and multiple agents.
• To evaluate the method's performance against state-of-the-art algorithms.

Main Methods:

• Utilized a homogeneous convolutional neural network for path generation.
• Applied the method to single path planning in 2D and 3D mazes.
• Extended the method to multi-path planning (single source to multiple endpoints, or vice versa).

Main Results:

• Achieved optimal or near-optimal paths in over 99.5% of single-agent cases (<10% longer).
• Successfully generated optimal or near-optimal paths for multi-agent scenarios (96.4% for two paths, 83.9% for three paths).
• Demonstrated competitive performance compared to existing state-of-the-art algorithms.

Conclusions:

• The proposed convolutional neural network method offers a highly effective and efficient solution for robot path planning.
• The one-shot approach significantly improves scalability for multi-agent systems.
• This novel method shows great promise for advancing autonomous robot navigation and manipulation.