New system uses semantic topology to help robots navigate cluttered indoor spaces

GIST (Grounded Intelligent Semantic Topology) is a research system designed to address spatial navigation challenges for embodied AI in complex indoor environments. The pipeline converts consumer-grade mobile point cloud data into a semantically annotated navigation topology, combining traditional topological mapping with vision-language model outputs.

The core architecture operates in stages: converting 3D point clouds into 2D occupancy maps, extracting topological structure, then overlaying semantic annotations through selective keyframe processing. This modular design allows the system to support multiple downstream tasks without requiring retraining for each application.

The researchers demonstrated four specific capabilities derived from the shared semantic topology: an intent-driven semantic search function that infers categorical alternatives when exact item matches fail; a semantic localizer reporting 1.04 m top-5 mean translation error; a zone classifier that segments walkable floor plans into semantic regions; and a natural language instruction generator that produces landmark-based egocentric routing cues.

A formative user evaluation with five participants (N=5) achieved an 80% navigation success rate when users relied solely on verbal instructions generated by the system. In LLM-based evaluations, the instruction generation component outperformed baseline sequence models, though the evaluation methodology and baseline details remain within the full paper.

The submission, authored by Shivendra Agrawal and Bradley Hayes, was posted to arXiv on April 16, 2026. The work spans computer vision, robotics, and human-computer interaction domains.