'SurfaceEncoder: Semi-Supervised Representation Learning for Global Wheelchair Accessibility'

For the millions of people who use wheelchairs worldwide, everyday sidewalks, curb ramps, and cobblestones imply great hardships and safety risks, yet mainstream navigation apps ignore them entirely! SurfaceEncoder tackles this with a semi-supervised deep learning framework that turns ordinary smartphone vibration data into accurate, automated maps of surface quality, slashing the need for the painstaking manual labeling that has blocked this work from scaling globally. We built the Global Wheelchair Surface (GWS) dataset: 183 hours and 52 million sensor points spanning five countries, pairing controlled labeled data from the USA and Vietnam with 300 km of unlabeled European city streets. The trick is that learning the physics of vibration from a little labeled data lets the model generalize across continents, boosting clustering quality by over 32% on unseen European surfaces and outperforming major time-series foundation models like TS2Vec, Chronos, MOMENT, and TimesFM. This isn't just a benchmark win: the pipeline powers a live, LLM-driven turn-by-turn navigation system that gives wheelchair users surface-aware directions in plain language. A huge thank-you to the reviewers for recognizing the social relevance, technical rigor, and the fact that we deployed a real working system. A small win for our AI for Social Good endeavor APPCAIR (ECML PKDD 2026 (CORE A) -- w/ valued collaborator, Prof. Vaskar Raychoudhury