Reproduce “PILOT: One Physics-Integrated Generation Framework to Unify 2D and 3D Radio Map Construction” on OpenResearch

About this paper

Weiming Huang, Hao Sun, Junting Chen

Researchers from The Chinese University of Hong Kong, Shenzhen, and City University of Hong Kong introduced PILOT, an autoregressive framework that constructs 2D and 3D pathloss radio maps by integrating physical propagation principles. This method achieves faster inference and improved accuracy over existing diffusion models while enabling zero-shot transfer across different carrier frequencies and environments.

Key insights

• Integrating physics-guided generation orders significantly reduces conditional uncertainty in autoregressive prediction, leading to a monotonically increasing entropy profile and improved accuracy, especially in complex environments.
• A refined tokenizer design, incorporating a linear projector, differentiable quantization, and a multi-scale 3D gradient regularizer, is critical for preventing codebook collapse and ensuring vertical continuity in 3D maps.
• A pathloss anchor map, combined with a robust VQ latent space, effectively anchors pathloss ranges and enables consistent zero-shot transfer across different carrier frequencies and scene geometries without retraining.

Summary by alphaXiv.