A developer has spent two years exploring AI techniques for autorouting in circuit board design, tackling the fundamental challenge of how to represent routing paths for machine learning models. Traditional grid-based approaches struggle with spatial understanding, while vision models excel at classification tasks. This led to exploring whether paths could be represented as patterns rather than grid coordinates.

The core insight involves training models to output paths as classifications of pattern sets rather than pixel-by-pixel decisions. This approach leverages the strengths of vision models while addressing the spatial reasoning limitations that have hampered AI-based autorouting systems. By representing paths as collections of patterns, the system can potentially capture the complex spatial relationships inherent in circuit board routing.

This recursive pattern-based approach represents a novel direction in autorouting AI research. The technique could bridge the gap between traditional algorithmic routing and modern machine learning approaches, potentially enabling more sophisticated and adaptable routing solutions for increasingly complex circuit designs.