Towards a Digital Twin Framework for Sensory-Aware Neurodivergent Route Personalisation

Around 90% of autistic adults report atypical sensory processing, which is a known barrier to independent travel. Existing navigation systems optimise for time and distance and do not account for the sensory demands of a route. We present NavTwin, a framework that pairs a Personal Digital Twin (PDT), which models a user’s sensitivity profile and preference history, with an Environmental Digital Twin (EDT), which uses a Temporal Graph Network to predict crowd density, noise, and visual complexity along a street graph. We use digital twin in a restricted sense: both twins are state-estimation and prediction models, and the EDT is not yet synchronised with live city feeds, so the loop back to the physical environment remains open. A route scorer combines personal preference, environmental comfort, completion probability, and efficiency, with weights adapted online by a Contextual Bandit. Route explanations are produced by a Gemma 3 1B Small Language Model fine-tuned with Retrieval Augmented Fine-Tuning and QLoRA, so that recommendations are grounded in curated domain documents rather than the model’s internal knowledge. Across four synthetic profiles, our adaptive scheme increases simulated journey completion for the high sensitivity profile by 9.6 percentage points over the strongest static baseline, with the bandit converging in 12–19 journeys. The evaluation is architecture-level, showing that the framework’s components behave coherently and lay the grounds for follow-on user studies with neurodivergent participants.