LATNAV: A Latency-Aware Real-Time Navigation Framework for Robots in Dynamic Human Environments

In human-populated environments, autonomous service robots are required to navigate safely and efficiently while adhering to stringent real-time constraints. Although recent advancements in deep learning and sensor fusion have significantly enhanced navigation accuracy, many systems regard computational latency as a secondary engineering consideration rather than a primary design variable. This study presents LATNAV (latency-aware transformer navigation), a comprehensive framework that integrates multimodal perception, short-horizon motion forecasting, anytime hierarchical planning, and safety-guaranteed control within a defined timing budget. The principal contribution is a latency-aware scheduler that assesses perception confidence, computational load, and deadline proximity to dynamically adjust model width, planner search depth, and control horizons at each cycle. LATNAV was evaluated on three open-access navigation benchmarks under varying crowd densities, sensor dropout conditions, and embedded compute constraints. The experimental findings indicate that LATNAV achieves a navigation success rate of 91.2% in medium-density scenarios (10 dynamic agents) and 82.5% in high-density scenarios (20 agents) while maintaining a deadline miss rate below 1.5%, representing a 10.8% relative improvement in success rate over the strongest non-adaptive baseline (TEB, 82.3%) and a 6.1× reduction in deadline violations compared with non-adaptive baselines. Ablation studies confirm that each component of the framework significantly contributes to both navigation quality and timing stability. Our results demonstrate that considering latency as a measurable and benchmarkable scientific output enhances the robustness and deployability of robot navigation in real-world human environments.