Multi-Sensor Fusion in Robotics: Core Frameworks, Practical Implementations and Development Trends

Abstract

As a core technology for enhancing environmental perception and autonomous decision-making in robots, multi-sensor fusion has advanced significantly in both theoretical and engineering aspects from 2015 to 2025. This paper reviews research progress through bibliometrics and case studies, identifying a three-stage evolution in algorithms: from Kalman Filter (KF)-based linear systems achieving around 40% accuracy improvement, to Particle Filter (PF) handling non-Gaussian noise with 85% localization accuracy, to deep learning-enabled end-to-end fusion reaching 92.7% classification accuracy. It analyzes common vision-LiDAR-IMU (Inertial Measurement Unit) configurations and compares traditional methods like EKF (Extended Kalman Filter) and UKF (Unscented Kalman Filter) with newer approaches such as CNN (Convolutional Neural Network) and GNN (Convolutional Neural Network). For example, LiDAR-vision SLAM (Simultaneous Localization and Mapping) attains 0.12 m RMSE (Root Mean Square Error). In applications including industrial robots, service robots, and autonomous driving, multi-sensor fusion improves localization accuracy to ±0.05 mm, dynamic obstacle avoidance success to 97%, and reduces mapping error to 2.3%. The study also addresses ongoing challenges like spatiotemporal registration, heterogeneous data fusion, and computational efficiency, while highlighting emerging trends such as federated learning and spiking neural networks. It offers theoretical and practical guidance for algorithm selection and optimization in multi-sensor robot systems.