Prediction-in-the-Loop Cyber-Physical Systems for Human-Aware Robot Navigation

Mixed human-robot environments on industrial floors demand navigation support decisions that are both safe and responsive under uncertainty. Reactive obstacle avoidance mitigates immediate collisions but underperforms in shared workspaces where human motion changes rapidly, and sensing is noisy or delayed. This paper presents a cyber-physical system (CPS) design centered on prediction-in-the-loop navigation, in which a neural trajectory forecaster continuously feeds short-horizon predictions of human motion into robot path planning. We frame the system as a feedback loop supporting runtime adaptation. The sensing layer employs ultra-wideband (UWB) indoor tracking, which is widely used for industrial localization and can reach decimeter-level typical errors in favorable conditions. Building on a lightweight recurrent predictor suitable for low-rate, privacy-preserving non-visual tracking, we provide a full methodology and system design for data ingestion, preprocessing, model training, and deployment as a containerized prediction service. Empirically, we quantify i) how preprocessing and early stopping influence prediction accuracy and mitigate instability in multi-step forecasts, ii) that multi-step prediction can be computed within a real-time control budget, and iii) that model caching and warm-start prediction services substantially improve end-to-end response time (RTT), reducing round-trip time from 1.0s to 0.18s in our tested configuration. We further specify a horizon-aware planning interface that converts prediction and sensing uncertainty into conservative safety envelopes suitable for downstream navigation controllers. The contributions are a reusable CPS architecture pattern for integrating learning-based predictors into runtime navigation loops, an end-to-end toolchain for UWB-to-prediction-to-planning integration, and an evaluation that characterizes key accuracy-latency trade-offs relevant to design automation for CPS and IoT.