Sensor Fusion-Based Navigation Systems for Autonomous Delivery Robots

Abstract

Autonomous delivery robots must navigate sidewalks, corridors, and mixed indoor–outdoor campuses while maintaining accuracy, safety, and efficiency under imperfect sensing. Single-sensor pipelines (e.g., wheel odometry or vision alone) degrade under wheel slip, poor lighting, occlusions, and multipath. This manuscript presents a sensor-fusion navigation architecture that integrates inertial measurement units (IMUs), wheel encoders, cameras, LiDAR, and optional ultra-wideband (UWB) anchors to achieve robust localization and motion planning in dynamic environments. We detail a modular stack: (1) time-synchronized preprocessing and calibration, (2) multi-rate odometry (wheel–IMU EKF, visual–inertial odometry, LiDAR–inertial odometry), (3) factor-graph smoothing with loop closures and UWB priors, (4) semantic mapping that separates static structure from dynamic obstacles, (5) dual-horizon planning with D*-Lite globally and model-predictive control (MPC) locally, and (6) a safety supervisor enforcing stop/slowdown under uncertainty spikes.

A simulation campaign across campus-sidewalk and urban-alley scenes (varying lighting, ground friction, and pedestrian density) compares four configurations: baseline wheel–IMU EKF, VIO-aided EKF, LiDAR-inertial odometry, and full multimodal factor-graph fusion including UWB. The fused system reduces absolute trajectory error by ~82% and collision rate by ~89% relative to the baseline, while adding <16 ms average fusion latency. We report statistically significant gains (ANOVA, Tukey HSD, p < 0.01) in success rate, path efficiency, and energy per kilometer. Results suggest that tightly-coupled, uncertainty-aware fusion—combined with semantic dynamics handling—yields navigation resilience suitable for last-meter delivery. We conclude with deployment guidance and open problems in long-term calibration drift, low-texture scenes, and learning-enhanced fusion.