Visible Light Positioning for Robotic Navigation

Abstract

Indoor positioning remains a persistent challenge where Global Positioning System (GPS) signals are unreliable. Visible Light Positioning (VLP) has emerged as a promising solution, leveraging the ubiquity of Light Emitting Diode (LED) lighting and low-cost photodiode receivers to provide centimeter-level accuracy. Despite its potential, two key barriers have limited progress: the absence of large-scale, high-quality datasets for algorithm development and evaluation, and the lack of evidence demonstrating that VLP can support practical robotic navigation.

This thesis addresses both challenges through a set of proposed frameworks supported by experimental investigation. The first study developed an autonomous ground-truthing and dataset collection system that co-locates a photodiode receiver with an HTC Vive tracker on a mobile robotic platform. This system enabled efficient acquisition of 7,344 fingerprint points within a controlled testbed, far exceeding the photodiode-based VLP RSS fingerprint datasets identified in the surveyed literature at the time of publication. Using this dataset, a comprehensive benchmarking of machine learning and channel model approaches for VLP was performed. Results showed that model-based methods retain an advantage when calibration data is scarce, but that machine learning techniques surpass them once trained on sufficiently large datasets. These findings resolved inconsistencies in the literature and highlighted the central role of data scale in VLP performance.

The second study demonstrated single-mode VLP navigation on a differential-drive robot under robust trajectory-level testing, then extended the investigation by combining VLP with odometry through a Long Short-Term Memory (LSTM)-based fusion model. Experimental evaluation showed that the fused approach maintains stable trajectories by learning sequential dependencies in the data. This demonstrated that VLP, when fused with odometry, can provide reliable localization and heading estimation for real robots operating in motion.

Together, these studies advance VLP research from controlled, small-scale laboratory investigations toward practical deployment in robotics. The thesis makes five key contributions: (i) a novel, automated ground-truth and data collection system for VLP, producing the largest experimental RSS fingerprint dataset for photodiode-based VLP identified in the surveyed literature at the time of publication, (ii) robust benchmarking of machine learning and model-based positioning methods, (iii) a real-world robotic system enabling photodiode-based VLP navigation with full pose estimation under ground truth, (iv) a comparative study of single-mode baselines versus LSTM-based fusion across offline and live navigation experiments, and (v) advancement of VLP from theoretical exploration toward practical deployment in robotic navigation.

While limitations remain, including sensitivity to orientation and tilt, confinement to controlled environments, and reliance on research-grade ground-truth systems, these highlight clear avenues for future work. Potential directions include benchmarking learning-based fusion against model-based filters (e.g., Extended Kalman Filter (EKF) variants), multi-photodiode or hybrid VLP–Inertial Measurement Unit (IMU) receivers, validation in large-scale industrial environments, extension to diverse robotic platforms, and integration with other positioning technologies.

In conclusion, this thesis demonstrates that VLP is a viable localization modality for robotic navigation. By overcoming barriers of dataset scale and by validating fusion-based navigation, the research provides a foundation for deploying VLP in real-world robotic and smart-building applications.