Abstract

In complex 3D game environments, players rely on visual affordances to spot critical map transition points. Efficient identification of such points is vital for client-side auto-mapping and offers an objective basis for evaluating cue presentation. We formalize the task of detecting traversable Spatial Transition Points (STPs)—connectors between two sub-regions—and selecting the singular Main STP (MSTP) that lies on the designer-intended critical path, from a single frame. We introduce a two-stage deep-learning pipeline that first detects potential STPs using Faster R-CNN and then ranks them with a lightweight MSTP selector fusing local and global visual features. Both stages leverage parameter-efficient adapters, with an optional retrieval-augmented fusion step. Experiments on a custom dataset from five action-RPG titles reveal that full-network fine-tuning excels when data are ample, while adapter-only transfer proves more robust in low-data scenarios and for MSTP selection. By defining this novel problem, providing a baseline pipeline and dataset, and highlighting trade-offs in model adaptation, we aim to spur future AI-driven navigation aids and data-informed level-design tools.

Citation

Xu, K., & Verbrugge, C. (2025, November). Adaptive Visual Navigation Assistant in 3D RPGs. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE 2025).

@inproceedings{xu2025navigation,
  author    = {Kaijie Xu and Clark Verbrugge},
  title     = {Adaptive Visual Navigation Assistant in 3D RPGs},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE)},
  year      = {2025},
  month     = {Nov}
}