Efficient Visibility Approximation for Game AI using Neural Omnidirectional Distance Fields

TL;DR

This work tackles the costly visibility tests in game AI by introducing neural Omnidirectional Distance Fields (ODFs) over partitioned scene regions. By mapping omnidirectional distance data onto a UV plane and encoding directions with a multi-resolution grid, a compact MLP can efficiently reconstruct directional distances to determine visibility without raycasting. The method achieves substantial runtime speedups and predictable inference time across multiple scenes, with offline accuracy around 0.90 and memory efficient representations compared to uncompressed depth maps. The approach enables scalable, CPU-friendly visibility queries suitable for real-time NPC behaviors, while recognizing areas for future enhancement such as dynamic scenes and edge-aware hybrids.

Abstract

Visibility information is critical in game AI applications, but the computational cost of raycasting-based methods poses a challenge for real-time systems. To address this challenge, we propose a novel method that represents a partitioned game scene as neural Omnidirectional Distance Fields (ODFs), allowing scalable and efficient visibility approximation between positions without raycasting. For each position of interest, we map its omnidirectional distance data from the spherical surface onto a UV plane. We then use multi-resolution grids and bilinearly interpolated features to encode directions. This allows us to use a compact multi-layer perceptron (MLP) to reconstruct the high-frequency directional distance data at these positions, ensuring fast inference speed. We demonstrate the effectiveness of our method through offline experiments and in-game evaluation. For in-game evaluation, we conduct a side-by-side comparison with raycasting-based visibility tests in three different scenes. Using a compact MLP (128 neurons and 2 layers), our method achieves an average cold start speedup of 9.35 times and warm start speedup of 4.8 times across these scenes. In addition, unlike the raycasting-based method, whose evaluation time is affected by the characteristics of the scenes, our method's evaluation time remains constant.

Figures (8)

• Figure 1: Omnidirectional distance data collection at a source position. (a) Directions are distributed on the spherical surface by a Fibonacci lattice, this visualization shows the total number of directions $P = 201$. (b) Collected omnidirectional distance data at one position of our offline experiment scene, with $P = 40001$.
• Figure 2: The effect of the high-frequency expressiveness of the neural network due to direction encoding in ODF reconstruction. As the truncated SH degree for direction encoding increases, the reconstructed distance data from the trained neural network contains more details.
• Figure 3: Illustration of multi-resolution grid encoding. (a) Spherical view of a two-level resolution grid in the longitude-latitude projection, where the sample direction is represented as "$\mathbf{x}$". (b) The (partial) UV plane view with the same grid and direction. The four nearest texels that store relevant features are highlighted for each grid.
• Figure 4: Offline evaluation environment.
• Figure 5: Visibility test throughput (kilo tests per second) of our method (PE, Long-Lat) for different batch sizes, tested on CPU and GPU using PyTorch.