Robust joint optical delay and received signal strength positioning for visible light communication systems using particle swarm optimization Authors: Iván Sánchez, Pablo Palacios Játiva, María Camila Reyes, David Zabala-Blanco, Cesar Azurdia Meza, Ismael Soto Published: 2026 ABSTRACT Accurate indoor positioning in Global Navigation Satellite Systems (GNSS)-denied environments remains a critical challenge for next-generation intelligent spaces, industrial automation, and context-aware services. Visible light communication (VLC) and visible light positioning (VLP) have emerged as attractive solutions due to their dual functionality of illumination and communication, immunity to electromagnetic interference, and potential for high-accuracy localization. However, practical VLC/VLP systems are highly sensitive to receiver orientation, LED/front-end nonlinearities, ambient light, multipath reflections, and hardware-related timing offsets and RSS measurement distortions. In this paper, we propose a robust joint optical delay–received signal strength (RSS) positioning framework for VLC systems, where delay-derived pseudorange information and received optical power measurements are fused under a unified weighted negative log-likelihood formulation. Unlike conventional VLP schemes that rely on either RSS-only or pseudorange-only inference under Gaussian assumptions, the proposed approach explicitly accounts for non-Gaussian measurement distortions and per-link timing bias in the optical domain. The user position and per-link timing-bias parameters are jointly estimated through particle swarm optimization (PSO), enabling robust operation in highly nonlinear and multimodal VLC localization landscapes. In the evaluated LOS-dominant scenario, the proposed joint optical delay–RSS estimator achieves an RMSE of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.24\,\textrm{m}$$\end{document} and a 95th percentile error of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.46\,\textrm{m}$$\end{document}, outperforming the Optical pseudorange WNLS, RSS-only Lambertian WNLS, and Joint pseudorange–RSS (PSO + Huber) baselines. This framework provides a calibration-efficient and Lambertian-consistent solution for indoor optical positioning and constitutes a promising basis for accurate localization in smart buildings, hospitals, warehouses, and industrial environments. These results should be interpreted within the simulation-based scope of the present study, and experimental validation remains necessary before drawing stronger deployment-oriented conclusions. Introduction Accurate indoor localization in GNSS-denied environments is important for asset tracking, automation, and context-aware services. Visible light communication (VLC) and visible light positioning (VLP) are attractive because deployed LED luminaires can simultaneously support illumination, communication, and localization, especially in environments where RF emissions are constrained1–3. VLC-based positioning benefits from directional optical propagation and reduced electromagnetic interference, but practical accuracy is strongly affected by receiver orientation, ambient light, partial shadowing, device imperfections, and non-line-of-sight (NLOS) reflections4,5. In addition, recent device-level studies have shown that GaN MQW LEDs exhibit injection-dependent nonlinear behavior in both illumination intensity and electro-optic response, which is not fully captured by a purely geometric Lambertian propagation model6. As a result, the relationship between measured observables and user position is highly nonlinear and often deviates from ideal Gaussian assumptions7–10. Most existing VLP techniques rely on a dominant observable, such as RSS, image-based measurements, timing-related features or machine learning algorithms11. RSS-based methods exploit LOS channel structure but are sensitive to orientation mismatch, gain uncertainty, and multipath distortion, whereas timing-based methods provide complementary geometric information but are affected by synchronization uncertainty, front-end delays, and measurement bias. Recent implementations include photodiode-array, OCC-based, and learning-driven approaches12–19. PSO-based optimization has also been explored in application-specific VLP settings, including RSSI-based positioning with random-angle interference cancellation and VLC localization in hospital environments5,20. A promising way to address these limitations is to combine heterogeneous observables within a joint estimation framework. Related multi-sensor and multi-observable studies in indoor positioning have shown clear robustness gains when complementary measurements are fused and bias terms are explicitly modeled21–24. Inspired by this idea, we argue that a similar fusion strategy can be reformulated in the optical domain, where timing-related delay information and optical RSS measurements can be jointly exploited to improve positioning robustness and accuracy. However, this adaptation is not a direct variable substitution. In VLC systems, the RSS observable must be derived from the Lambertian optical channel model, while the timing-related observable must account for optical delay estimation mechanisms and hardware-dependent bias terms. At the same time, it is important to distinguish the propagation model from the electro-optic device model: the present paper focuses on positioning observables and LOS propagation geometry, whereas explicit injection-dependent LED nonlinearity is not modeled through a compact MQW device description. In addition, the error statistics in VLC are often asymmetric or heavy-tailed due to clipping, saturation, ambient illumination, and intermittent LOS blockage, which makes Gaussian-based formulations insufficient. Motivated by these challenges, this paper proposes a robust joint optical delay–RSS positioning framework for VLC systems. The main idea is to combine delay-derived pseudorange information and optical power measurements in a unified weighted negative log-likelihood formulation, while explicitly accounting for non-Gaussian residuals and per-link timing bias. The resulting optimization problem is highly nonlinear and may contain multiple local minima due to the geometry of the optical channel and the coupling between position, incidence angle, and received power. To address this issue, we adopt particle swarm optimization (PSO) as a global search strategy capable of exploring multimodal cost surfaces without requiring gradient information. This makes the resulting estimator well suited for practical VLC localization scenarios, especially when coarse initialization is available but model mismatch and measurement irregularities remain significant25,26. The novelty claimed here is therefore not the isolated use of PSO, but its integration with a joint optical delay–RSS weighted negative log-likelihood, explicit per-link bias estimation, and robust heavy-tailed residual modeling within a unified VLC positioning framework. It should be emphasized that the present study is methodological and simulation-based. In particular, the optical-delay observable is introduced as a synchronization-assisted pseudorange abstraction to assess the benefit of robust delay–RSS fusion under realistic bias and non-Gaussian disturbances, rather than as an immediately deployable hardware prototype. This joint formulation is attractive because it combines complementary delay and RSS observables, absorbs per-link timing offsets through explicit bias terms, and uses robust likelihood modeling to tolerate asymmetric and non-Gaussian disturbances. In principle, it is also extensible to three-dimensional positioning, receiver orientation estimation, angle-diversity reception, and hybrid LOS/NLOS conditions9,22. The main contributions of this paper are summarized as follows:We propose a joint optical delay–RSS positioning framework for VLC systems that fuses timing-based and power-based observables in a unified estimation model.We formulate the localization problem using a weighted negative log-likelihood cost function that explicitly accommodates non-Gaussian residual behavior and per-link bias in the optical domain.We use PSO as the optimization engine for the resulting nonlinear and multimodal localization problem, while the main contribution remains the proposed joint delay–RSS, bias-aware, and robust estimator formulation.We provide a VLC-oriented formulation that is consistent with the adopted Lambertian LOS propagation model and suitable for LOS-dominant indoor positioning scenarios. We propose a joint optical delay–RSS positioning framework for VLC systems that fuses timing-based and power-based observables in a unified estimation model. We propose a joint optical delay–RSS positioning framework for VLC systems that fuses timing-based and power-based observables in a unified estimation model. We formulate the localization problem using a weighted negative log-likelihood cost function that explicitly accommodates non-Gaussian residual behavior and per-link bias in the optical domain. We formulate the localization problem using a weighted negative log-likelihood cost function that explicitly accommodates non-Gaussian residual behavior and per-link bias in the optical domain. We use PSO as the optimization engine for the resulting nonlinear and multimodal localization problem, while the main contribution remains the proposed joint delay–RSS, bias-aware, and robust estimator formulation. We use PSO as the optimization engine for the resulting nonlinear and multimodal localization problem, while the main contribution remains the proposed joint delay–RSS, bias-aware, and robust estimator formulation. We provide a VLC-oriented formulation that is consistent with the adopted Lambertian LOS propagation model and suitable for LOS-dominant indoor positioning scenarios. We provide a VLC-oriented formulation that is consistent with the adopted Lambertian LOS propagation model and suitable for LOS-dominant indoor positioning scenarios. Accordingly, the contribution of this work should be understood at the level of the complete estimator: not the isolated use of PSO, RSS, or multimodal fusion, but the integration of a VLC-specific optical delay–RSS observation model, explicit per-link bias estimation, and robust heavy-tailed likelihood design within a single positioning framework. The remainder of this paper is organized as follows. The Methods section presents the VLC system model and the proposed joint optical delay–RSS estimator. The Results section introduces the simulation setup and reports comparative localization performance. The Discussion section interprets the main findings, practical implications, and limitations. Finally, the Conclusion summarizes the contributions and outlines future research directions. Methods System model In this section, we introduce the considered indoor VLC positioning scenario and the mathematical models associated with the two observables exploited in this work, namely optical delay and RSS. The framework considered here assumes a set of LED luminaires with known locations and a mobile user equipped with a photodetector (PD) receiver whose position is to be estimated. Indoor VLC positioning scenario Let us consider an indoor VLP environment with N LED transmitters deployed at known coordinates1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{l}_i = [x_i,\, y_i,\, z_i]^T,\quad i=1,2,\dots ,N, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{l}_i \in \mathbb {R}^3$$\end{document} denotes the position of the i-th LED luminaire. The VLC receiver, equipped with a photodetector, is located at an unknown position2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{p} = [x,\, y,\, z]^T. \end{aligned}$$\end{document}The Euclidean distance between the i-th LED and the receiver is given by3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d_i(\textbf{p}) = \Vert \textbf{p} - \textbf{l}_i\Vert _2. \end{aligned}$$\end{document}We assume that the coordinates of the luminaires are known a priori through infrastructure planning or system calibration, while the receiver position is the target of estimation. Each LED periodically transmits a known optical pilot or beacon sequence that enables the extraction of timing-related and power-related observables at the receiver side. The receiver is assumed to be able to identify the transmitting LED associated with each received beacon. The proposed model is general and can be applied to three-dimensional localization. However, if the receiver height is known or approximately fixed, the positioning problem can be simplified to a two-dimensional horizontal localization task. In the most general case, the receiver coordinates are jointly estimated in three dimensions. Optical channel geometry The optical wireless channel in VLC is primarily determined by the relative geometry between the LED transmitter and the PD receiver. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi _i$$\end{document} denote the irradiance angle of the i-th LED, and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi _i$$\end{document} denote the angle of incidence at the receiver. These angles depend on the orientations of the transmitter and receiver as well as on their relative positions. Assuming downward-facing LEDs and an upward-facing receiver for the nominal LOS configuration, the channel geometry can be described using the vector connecting the LED and the receiver:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{r}_i = \textbf{p} - \textbf{l}_i. \end{aligned}$$\end{document}Then, the propagation distance is5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d_i = \Vert \textbf{r}_i\Vert _2. \end{aligned}$$\end{document}Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{n}_{t,i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{n}_r$$\end{document} denote the unit orientation vectors of the i-th LED transmitter and the receiver, respectively. The irradiance and incidence angles can be written as6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \cos (\phi _i) = \frac{\textbf{n}_{t,i}^T (\textbf{p} - \textbf{l}_i)}{\Vert \textbf{p} - \textbf{l}_i\Vert _2}, \end{aligned}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \cos (\psi _i) = \frac{\textbf{n}_r^T (\textbf{l}_i - \textbf{p})}{\Vert \textbf{l}_i - \textbf{p}\Vert _2}. \end{aligned}$$\end{document}The received optical signal is only considered valid if the receiver is within the FOV, that is,8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} 0 \le \psi _i \le \Psi _c, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Psi _c$$\end{document} is the receiver FOV semi-angle. RSS-based observation model Under LOS conditions, the direct-current (DC) channel gain of the optical link from the i-th LED to the receiver can be expressed using the Lambertian radiation model as59\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} H_i(\textbf{p})&= \frac{(m_i+1)A_r}{2\pi d_i^2} \cos ^{m_i}(\phi _i)\, T_s(\psi _i)\\&\quad \times g(\psi _i)\, \cos (\psi _i), \end{aligned} \end{aligned}$$\end{document}for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0 \le \psi _i \le \Psi _c$$\end{document}, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_i(\textbf{p})=0$$\end{document} when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi _i> \Psi _c$$\end{document}. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m_i$$\end{document} is the Lambertian order of the i-th LED, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_r$$\end{document} is the active area of the PD, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_s(\psi _i)$$\end{document} is the gain of the optical filter, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g(\psi _i)$$\end{document} is the concentrator gain, typically modeled as10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g(\psi _i) = {\left\{ \begin{array}{ll} \displaystyle \frac{n_c^2}{\sin ^2(\Psi _c)}, & 0 \le \psi _i \le \Psi _c,\\ 0, & \psi _i> \Psi _c, \end{array}\right. } \end{aligned}$$\end{document}with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_c$$\end{document} denoting the refractive index of the optical concentrator. The Lambertian order is related to the LED semi-angle at half power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi _{1/2}$$\end{document} through511\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} m_i = -\frac{\ln 2}{\ln \left( \cos \Phi _{1/2}\right) }. \end{aligned}$$\end{document}The received optical power associated with the i-th LED can then be written as512\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P_{r,i}(\textbf{p}) = P_{t,i} H_i(\textbf{p}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{t,i}$$\end{document} is the transmitted optical power of the i-th luminaire. It is important to emphasize that (12) is a propagation-level received-power model conditioned on the emitted optical power. It captures LOS geometric attenuation and angle dependence, but it does not explicitly represent the injection-dependent electro-optic conversion inside the LED. Recent physics-based compact modeling of GaN MQW LEDs has shown that both the emitted illumination intensity and the communication response vary nonlinearly with operating point6. Therefore, in the present work the Lambertian model should be interpreted as a system-level propagation model, while the residual terms introduced later act as aggregate observation-level surrogates for front-end distortion and measurement corruption. In practice, the observed RSS measurement is affected by ambient light, thermal noise, shot noise, front-end gain variation, clipping, and possible multipath distortion. Therefore, the measured optical RSS is modeled as13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{P}_{r,i} = P_{r,i}(\textbf{p}) + n_i^{(P)}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_i^{(P)}$$\end{document} represents the RSS measurement error. In realistic VLC scenarios, this residual term is not necessarily Gaussian, especially under partial blockage, orientation mismatch, or saturation effects. Optical delay-based observation model In addition to the RSS observable, timing-related information can be extracted from the received optical beacon sequence. Depending on the implementation, this observable may be obtained from ToA, synchronization-assisted delay estimation, or delay-derived pseudoranging. To maintain generality, we refer to this observable as optical delay, and denote its delay-derived pseudorange by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\tau }_i$$\end{document} denote the estimated propagation-related delay associated with the signal received from the i-th LED. The corresponding delay-derived pseudorange observation is written as14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\rho }_i = c\,\hat{\tau }_i, \end{aligned}$$\end{document}where c denotes the speed of light. The adopted delay-derived pseudorange observation model relating the measurement to the receiver position is15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\rho }_i = d_i(\textbf{p}) + b_i + n_i^{(\rho )}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document} is a bias term associated with hardware timing offsets, synchronization mismatch, or systematic front-end delay, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_i^{(\rho )}$$\end{document} denotes the random error affecting the delay estimate. The bias term is particularly relevant in practical optical positioning systems because the measured timing observable may include delays introduced by LED drivers, signal processing chains, detection circuitry, and synchronization imperfections. As in the RSS case, the delay residual may exhibit non-Gaussian behavior due to interference, threshold effects, imperfect beacon detection, or distorted pulse shapes. Joint observation vector For each LED transmitter, two complementary observables are therefore available: a power-domain measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document} and a delay-derived pseudorange measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}. Stacking all measurements, the overall observation set can be expressed as16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {Z} = \left\{ \hat{P}_{r,1}, \dots , \hat{P}_{r,N}, \hat{\rho }_1, \dots , \hat{\rho }_N \right\} . \end{aligned}$$\end{document}The objective of the positioning problem is to estimate the unknown user position \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{p}$$\end{document}, and optionally the bias vector17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{b} = [b_1, b_2, \dots , b_N]^T, \end{aligned}$$\end{document}from the heterogeneous observation set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {Z}$$\end{document}. The use of both observables is motivated by their complementary nature. On the one hand, delay-related measurements provide geometric ranging information that is directly linked to the propagation distance. On the other hand, RSS measurements provide additional spatial constraints through the optical channel gain, which depends not only on distance but also on incidence and irradiance angles. Jointly exploiting both observables is therefore expected to improve robustness and positioning accuracy compared to single-observable methods. Localization problem statement Based on the previous models, the localization task can be stated as the problem of estimating the receiver position from noisy and possibly biased optical measurements. Let18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\boldsymbol{\theta }} = \left\{ \begin{array}{ll} \textbf{p}, & \text {position-only estimation},\\ {[}{\textbf{p}}^{T},{\textbf{b}}^{T}]^T, & \text {joint position-and-bias estimation}. \end{array}\right. \end{aligned}$$\end{document}The estimation problem can then be expressed in general form as19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\hat{\boldsymbol{\theta }}} = \arg \min _{\boldsymbol{\theta }} \; \mathscr {C}({\boldsymbol{\theta }}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {C}(\boldsymbol{\theta })$$\end{document} is a cost function constructed from the mismatch between the measured observables and their model-based counterparts. Since both the RSS and delay models are nonlinear with respect to the receiver position, and because the residual statistics may deviate from ideal Gaussian assumptions, the resulting optimization landscape can be highly nonconvex and multimodal. This motivates the use of a robust joint likelihood-based formulation and a global optimization strategy, which will be developed in the next section. Proposed joint robust positioning method This section presents the proposed localization framework based on the joint exploitation of synchronization-assisted optical delay pseudoranges and received optical power measurements. First, the adopted estimation model is introduced. Then, a weighted robust negative log-likelihood cost function is formulated, followed by the PSO-based solution strategy. Optical delay pseudorange formulation In practical VLC systems, directly implementing round-trip time (RTT) estimation is not always the most convenient approach due to the stringent timing requirements imposed by the speed of light and the hardware complexity of bidirectional optical timing exchanges. Instead, a more suitable formulation consists of using synchronization-assisted one-way delay estimation, where each LED periodically transmits a known beacon and the receiver estimates the corresponding time of arrival relative to a common timing reference or a pre-established synchronization structure. At meter-scale indoor distances, the corresponding propagation times are on the order of only a few nanoseconds, so a direct experimental implementation would require tightly synchronized transmitters and receivers, high-bandwidth front ends, and specialized timestamping or time-to-digital conversion hardware. For reference, the decimeter-to-centimeter pseudorange disturbances used later in the numerical section correspond to sub-nanosecond residual timing errors; for example, a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.10\,\textrm{m}$$\end{document} bias corresponds to about \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.33\,\textrm{ns}$$\end{document}, while a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.06\,\textrm{m}$$\end{document} scale corresponds to about \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.20\,\textrm{ns}$$\end{document}. The exact beacon bandwidth, waveform, and sampling strategy are implementation dependent and are therefore not fixed in the present system-level study. Consequently, the adopted pseudorange errors should be interpreted as post-synchronization, post-processing residual timing errors rather than as raw sampling intervals. For this reason, the present paper focuses on the system-level estimation model and leaves full hardware realization of the optical-delay block for future work. For this reason, the timing-related observable considered in this work is modeled as an optical delay pseudorange. Specifically, for the i-th LED, the measured pseudorange is written as20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\rho }_i = c\,\hat{\tau }_i, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\tau }_i$$\end{document} is the estimated optical propagation-related delay and c is the speed of light. Unlike an ideal geometric range, the measured quantity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document} may include hardware- and synchronization-related offsets. Accordingly, the optical delay pseudorange model is expressed as21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\rho }_i = d_i(\textbf{p}) + b_i + n_i^{(\rho )}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i(\textbf{p})$$\end{document} is the Euclidean distance between the receiver and the i-th LED, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document} is a bias term accounting for systematic timing offsets, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_i^{(\rho )}$$\end{document} denotes the residual error. The use of per-link bias terms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document} is justified by the fact that different optical transmitters and front-end processing chains may introduce non-identical delays, especially in distributed VLC infrastructures. These biases may originate from driver circuitry, beacon generation latency, clock misalignment, signal thresholding, analog front-end delays, or implementation-dependent timestamping effects. Therefore, the pseudorange model in (21) provides a flexible and physically meaningful abstraction for timing-based VLC positioning. Joint observation model Let the unknown parameter vector be defined as22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \boldsymbol{\theta } = [\textbf{p}^T,\textbf{b}^T]^T = [x,\,y,\,z,\,b_1,\,b_2,\dots ,b_N]^T, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{p}=[x,y,z]^T$$\end{document} is the receiver position and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}=[b_1,b_2,\dots ,b_N]^T$$\end{document} is the bias vector. The 3D form is retained for generality, but in the numerical evaluation of this paper the receiver height is assumed known, so the effective estimated parameter vector becomes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[x,\,y,\,b_1,\dots ,b_N]^T$$\end{document}. For each LED transmitter, two heterogeneous observations are available:an optical delay pseudorange measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document},an RSS measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. an optical delay pseudorange measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}, an optical delay pseudorange measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}, an RSS measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. an RSS measurement \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. The model-based counterparts of these observations are23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho _i(\boldsymbol{\theta }) = d_i(\textbf{p}) + b_i, \end{aligned}$$\end{document}and24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P_{r,i}(\boldsymbol{\theta }) = P_{t,i} H_i(\textbf{p}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_i(\textbf{p})$$\end{document} is the VLC channel gain defined in the System Model subsection. The corresponding residuals are defined as25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} e_i^{(\rho )}(\boldsymbol{\theta }) = \hat{\rho }_i - \rho _i(\boldsymbol{\theta }) = \hat{\rho }_i - d_i(\textbf{p}) - b_i, \end{aligned}$$\end{document}and26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} e_i^{(P)}(\boldsymbol{\theta }) = \hat{P}_{r,i} - P_{r,i}(\boldsymbol{\theta }). \end{aligned}$$\end{document}These residuals reflect the mismatch between the measured optical observables and the position-dependent channel model. Since both observables are affected by nonlinear geometry and practical hardware impairments, the residuals generally exhibit non-Gaussian behavior. As a result, robust likelihood-based modeling is more appropriate than conventional least-squares formulations. Robust weighted negative log-likelihood formulation To jointly exploit the two observables, we formulate the localization problem as the minimization of a weighted negative log-likelihood (NLL) cost. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{\rho }(\cdot )$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{P}(\cdot )$$\end{document} denote the probability density functions associated with the pseudorange and RSS residuals, respectively. Without restricting the formulation to a specific residual distribution, the joint cost function is written as27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {C}(\boldsymbol{\theta }) = w_{\rho }\,\mathscr {C}_{\rho }(\boldsymbol{\theta }) + w_{P}\,\mathscr {C}_{P}(\boldsymbol{\theta }), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{\rho }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_P$$\end{document} are nonnegative weighting factors, and28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {C}_{\rho }(\boldsymbol{\theta }) = \sum _{i=1}^{N} -\log f_{\rho } \!\left( e_i^{(\rho )}(\boldsymbol{\theta }) \right) , \end{aligned}$$\end{document}29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {C}_{P}(\boldsymbol{\theta }) = \sum _{i=1}^{N} -\log f_{P} \!\left( e_i^{(P)}(\boldsymbol{\theta }) \right) . \end{aligned}$$\end{document}The weighting factors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{\rho }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_P$$\end{document} control the relative influence of the delay and RSS domains. In general, these weights can be selected according to prior measurement confidence, calibration data, estimated variances, or empirical normalization criteria. A practical choice is to define them inversely proportional to the nominal dispersion of each observable, thus preventing either domain from dominating the optimization solely due to scale mismatch. To make the formulation concrete while preserving robustness, we adopt a generalized heavy-tailed residual model. One suitable option is the Laplacian likelihood, which yields the following cost components:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {C}_{\rho }(\boldsymbol{\theta }) = \sum _{i=1}^{N} \frac{|e_i^{(\rho )}(\boldsymbol{\theta })|}{\beta _{\rho }}, \end{aligned}$$\end{document}31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathscr {C}_{P}(\boldsymbol{\theta }) = \sum _{i=1}^{N} \frac{|e_i^{(P)}(\boldsymbol{\theta })|}{\beta _{P}}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{\rho }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _P$$\end{document} are scale parameters associated with the pseudorange and RSS residuals, respectively. The adoption of Laplacian-type penalties is well justified in VLC localization because measurement errors may become heavy-tailed under partial LOS blockage, ambient illumination bursts, clipping, gain uncertainty, orientation mismatch, and imperfect threshold-based timing extraction. Compared with quadratic least-squares penalties, the proposed NLL formulation is less sensitive to occasional outliers and therefore better suited for practical optical positioning scenarios. Accordingly, the final optimization problem is expressed as32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\boldsymbol{\theta }} = \arg \min _{\boldsymbol{\theta }\in \Omega } \mathscr {C}(\boldsymbol{\theta }), \end{aligned}$$\end{document}where Ω denotes the feasible search region, defined by the room dimensions, admissible receiver height range, and physically plausible bias intervals. Identifiability considerations The identifiability of joint position-and-bias estimation depends on both the number of observables and their sensitivity to different subsets of the unknown parameters. In the specific setting evaluated in this paper, the receiver height is fixed and four LEDs are used, so the estimator processes eight scalar observations (four optical pseudoranges and four RSS values) to estimate six unknowns, namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[x,y,b_1,b_2,b_3,b_4]^T$$\end{document}. Under these conditions, the problem is not equivalent to unconstrained pseudorange-only trilateration with free per-link biases. If only the pseudorange equations were used, the bias terms could indeed absorb a significant part of the geometric mismatch, making the problem weakly constrained. However, the RSS equations do not depend on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document}; they constrain only the receiver position through the Lambertian channel geometry. Consequently, the bias variables cannot improve the pseudorange fit without remaining compatible with the RSS measurements. For nondegenerate anchor placement and locally informative RSS gradients, the combined delay–RSS model is therefore locally well constrained in the 2D fixed-height scenario considered here. In addition, the per-link bias vector is interpreted in this work as a quasi-static nuisance parameter over one estimation batch, not as a fast time-varying random process. Accordingly, the bias vector is re-estimated independently at each receiver location, rather than calibrated once over multiple known points. In this sense, the four-LED configuration should be interpreted as a minimum working setup for the fixed-height 2D case studied here, whereas additional anchors would improve redundancy and numerical robustness in noisier or more general deployments. More general cases, such as fully 3D positioning with unknown height, explicit orientation estimation, or rapidly time-varying per-link biases, would generally require extra observability support, for example through additional anchors, temporal regularization, or prior calibration constraints. Normalization and practical weight selection Since optical pseudorange residuals and RSS residuals are expressed in different units, a proper normalization strategy is required to ensure numerical balance in the joint objective. To this end, the weighting coefficients can be selected as33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} w_{\rho } = \frac{1}{\sigma _{\rho }+\epsilon }, \qquad w_{P} = \frac{1}{\sigma _{P}+\epsilon }, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{\rho }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _P$$\end{document} denote the nominal scales of the corresponding residuals and ε>0 is a small constant introduced to avoid numerical instability. Under the Laplacian penalties in (30)–(31), a convenient practical default is to let the scale parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{\rho }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _P$$\end{document} absorb most of the cross-domain normalization, in which case \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{\rho }=w_P=1$$\end{document} becomes a reasonable choice. If a short calibration batch is available, robust Laplacian-consistent scale estimates can be obtained as34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\beta }_{\rho } = \frac{\textrm{median}_j\left| r_j^{(\rho )}-\textrm{median}(\{r_j^{(\rho )}\})\right| }{\ln 2}, \qquad \hat{\beta }_{P} = \frac{\textrm{median}_j\left| r_j^{(P)}-\textrm{median}(\{r_j^{(P)}\})\right| }{\ln 2}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_j^{(\rho )}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_j^{(P)}$$\end{document} denote calibration residuals in the delay and RSS domains, respectively. These residuals may be obtained from a short startup calibration stage, from measurements at known reference points, or from a coarse preliminary estimator. In the numerical evaluation of this paper, the scale parameters are fixed to the nominal disturbance levels used in the synthetic measurement generator, namely \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{\rho }=0.06\,\textrm{m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _P=0.08\,P_{r,i}^{\textrm{max}}$$\end{document}, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{\rho }=w_P=1$$\end{document} is used throughout. Importantly, these quantities are kept fixed across all test points and compared methods rather than tuned per scenario, so that the reported gains are not driven by hyperparameter retuning. If calibration data are not available, (33) provides a conservative alternative based on nominal modality dispersions or sensor specifications. PSO-based estimation strategy The optimization problem in (32) is nonlinear and generally nonconvex because both the pseudorange model and the RSS channel gain depend on the receiver position in a highly nonlinear manner. Moreover, the joint inclusion of bias parameters increases the dimensionality of the search space and may introduce multiple local minima. For these reasons, gradient-based solvers may converge to suboptimal solutions, particularly when the initial guess is inaccurate or when the measurements are corrupted by non-Gaussian disturbances. To address this issue, we employ PSO as a derivative-free global search strategy. PSO is well suited for multimodal optimization problems and has the advantage of being simple to implement while requiring only repeated evaluations of the objective function. Let the swarm contain S particles. The state of the s-th particle at iteration k is denoted by35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{x}_s^{(k)} \in \mathbb {R}^{3+N}, \end{aligned}$$\end{document}with associated velocity vector36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{v}_s^{(k)} \in \mathbb {R}^{3+N}. \end{aligned}$$\end{document}Each particle represents a candidate solution for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\theta }$$\end{document}. The velocity update rule is given by37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} \textbf{v}_s^{(k+1)}&= \omega \textbf{v}_s^{(k)} + c_1 r_1^{(k)} \left( \textbf{p}_s^{\star } - \textbf{x}_s^{(k)} \right) \\&\quad + c_2 r_2^{(k)} \left( \textbf{g}^{\star } - \textbf{x}_s^{(k)} \right) , \end{aligned} \end{aligned}$$\end{document}where ω is the inertia factor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_1$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_2$$\end{document} are the cognitive and social acceleration coefficients, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_1^{(k)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_2^{(k)}$$\end{document} are independent random scalars uniformly distributed in [0,1], \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{p}_s^{\star }$$\end{document} is the personal best position found by particle s, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g}^{\star }$$\end{document} is the global best solution found by the swarm. The particle position is then updated as38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{x}_s^{(k+1)} = \textbf{x}_s^{(k)} + \textbf{v}_s^{(k+1)}. \end{aligned}$$\end{document}At each iteration, the objective value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {C}(\textbf{x}_s^{(k)})$$\end{document} is evaluated for all particles, and both personal and global best solutions are updated accordingly. Boundary constraints are enforced by projecting infeasible solutions back into the admissible region Ω. The algorithm stops when a maximum number of iterations is reached or when the improvement of the global best cost falls below a predefined threshold. The final estimate is given by the best particle found during the iterative search:39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\boldsymbol{\theta }} = \textbf{g}^{\star }. \end{aligned}$$\end{document} Algorithmic Summary The main steps of the proposed method are summarized as follows: Acquire the set of VLC measurements \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\hat{\rho }_i,\hat{P}_{r,i}\}_{i=1}^{N}$$\end{document} from synchronized optical beacons.Define the search space Ω according to room geometry and admissible bias intervals.Initialize a swarm of particles representing candidate vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\theta }$$\end{document}.Evaluate the robust weighted NLL cost in (27) for all particles.Update the particle velocities and positions using (37) and (38).Update personal and global best solutions.Repeat the previous two steps until convergence.Return the final receiver position estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{p}}$$\end{document} and the associated bias estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{b}}$$\end{document}. Acquire the set of VLC measurements \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\hat{\rho }_i,\hat{P}_{r,i}\}_{i=1}^{N}$$\end{document} from synchronized optical beacons. Acquire the set of VLC measurements \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{\hat{\rho }_i,\hat{P}_{r,i}\}_{i=1}^{N}$$\end{document} from synchronized optical beacons. Define the search space Ω according to room geometry and admissible bias intervals. Define the search space Ω according to room geometry and admissible bias intervals. Initialize a swarm of particles representing candidate vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\theta }$$\end{document}. Initialize a swarm of particles representing candidate vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\theta }$$\end{document}. Evaluate the robust weighted NLL cost in (27) for all particles. Evaluate the robust weighted NLL cost in (27) for all particles. Update the particle velocities and positions using (37) and (38). Update the particle velocities and positions using (37) and (38). Update personal and global best solutions. Update personal and global best solutions. Repeat the previous two steps until convergence. Repeat the previous two steps until convergence. Return the final receiver position estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{p}}$$\end{document} and the associated bias estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{b}}$$\end{document}. Return the final receiver position estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{p}}$$\end{document} and the associated bias estimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\textbf{b}}$$\end{document}. Methodological considerations The proposed formulation differs from conventional RSS-only VLP methods by incorporating an additional timing-based observable that provides direct geometric information. It also differs from purely range-based optical localization methods by introducing RSS as a complementary source of spatial discrimination through the Lambertian channel response. The resulting framework is particularly attractive in scenarios where one of the observables becomes temporarily less reliable. For example, RSS may be strongly perturbed by receiver tilt or gain uncertainty, whereas pseudorange measurements may remain informative; conversely, when timing estimates become noisy, the RSS observable can still constrain the search space. Furthermore, the explicit estimation of the bias vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}$$\end{document} enhances the practical relevance of the method, since it avoids the need for idealized assumptions regarding perfectly synchronized or delay-free optical hardware. This feature is especially important when considering real VLC implementations in which different luminaires may exhibit unequal processing latencies. Under the adopted system-level assumptions, the combination of robust residual modeling, heterogeneous optical observables, and PSO-based global optimization therefore provides a flexible estimation framework for indoor visible light positioning. Results In this section, the performance of the proposed joint optical delay–RSS positioning algorithm is evaluated in a representative indoor VLC scenario. We consider a controlled LOS-dominant environment, a finite set of reference test points, and a comparative analysis against baseline estimators using the cumulative distribution function (CDF) and standard localization metrics such as root mean square error (RMSE), 95th percentile error, and standard deviation. Simulation scenario We consider a rectangular indoor room of dimensions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8~\textrm{m} \times 11~\textrm{m} \times 3~\textrm{m}$$\end{document}. Four LED luminaires are mounted on the ceiling and act as optical anchors with known coordinates40\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{l}_1 = [1.0,\,1.0,\,3.0]^T,\quad \textbf{l}_2 = [7.0,\,1.0,\,3.0]^T, \end{aligned}$$\end{document}41\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{l}_3 = [1.0,\,10.0,\,3.0]^T,\quad \textbf{l}_4 = [7.0,\,10.0,\,3.0]^T. \end{aligned}$$\end{document}The receiver is assumed to be located at a height of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.85~\textrm{m}$$\end{document}, corresponding to a handheld or mobile terminal. The positioning task is therefore reduced to horizontal two-dimensional localization over the plane \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z=0.85~\textrm{m}$$\end{document}, while the LED heights remain fixed and known. The target positions are sampled over 48 reference points uniformly distributed over the accessible floor area. Fig. 1 depicts the top-view geometry of the indoor VLC positioning scenario considered in this work. Four ceiling-mounted LED luminaires are used as optical anchors, while 48 reference points are uniformly distributed over the accessible floor area to evaluate localization performance. This geometry provides a balanced but still practically relevant benchmark for assessing the behavior of the proposed estimator under LOS-dominant optical propagation.Fig. 1Top-view layout of the considered \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8~\textrm{m}\times 11~\textrm{m}$$\end{document} indoor VLC positioning scenario, showing the four ceiling-mounted LED anchors and the 48 reference test points used in the numerical evaluation. Top-view layout of the considered \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8~\textrm{m}\times 11~\textrm{m}$$\end{document} indoor VLC positioning scenario, showing the four ceiling-mounted LED anchors and the 48 reference test points used in the numerical evaluation. All luminaires are assumed to be downward-facing and identical, with semi-angle at half power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi _{1/2}=60^\circ$$\end{document}, transmitted optical power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t=20~\textrm{W}$$\end{document}, and direct-current optical modulation suitable for beacon-based synchronization and power estimation. The photodetector parameters are set as follows: active area \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_r=1~\textrm{cm}^2$$\end{document}, refractive index \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_c=1.5$$\end{document}, and field of view \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Psi _c=70^\circ$$\end{document}. For simplicity, the optical filter gain is set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_s(\psi )=1$$\end{document} within the FOV. The nominal receiver is assumed upward-facing in the main experiment, while later sensitivity analyses include random orientation perturbations. In all cases, the LOS path is preserved, although the effective RSS is affected by geometric angular dependence and by additive disturbances that emulate realistic VLC measurements. Measurement generation For each reference point and each optical anchor, two measurements are generated: an optical delay pseudorange observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document},an RSS observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. an optical delay pseudorange observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}, an optical delay pseudorange observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{\rho }_i$$\end{document}, an RSS observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. an RSS observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{P}_{r,i}$$\end{document}. The noiseless optical RSS is computed from the Lambertian LOS channel model introduced in the Methods section. The optical delay pseudorange is generated from the true geometric distance plus a link-dependent bias term. Specifically, the pseudorange model is42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\rho }_i = d_i(\textbf{p}) + b_i + n_i^{(\rho )}, \end{aligned}$$\end{document}where the bias term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_i$$\end{document} is randomly drawn for each LED from the interval43\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} b_i \sim \mathscr {U}(-0.10,\,0.10)\;\textrm{m}, \end{aligned}$$\end{document}which emulates systematic optical timing offsets due to driver and front-end processing delays. For each test point, a single bias vector is drawn once and then kept fixed over the corresponding M=50 measurement snapshots, while the residual noise and occasional outliers vary from snapshot to snapshot. This matches the intended interpretation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}$$\end{document} as a slowly varying link-dependent hardware offset rather than a per-snapshot random disturbance. The pseudorange residual \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_i^{(\rho )}$$\end{document} is modeled as a heavy-tailed random term in order to emulate practical beacon timing imperfections. In particular, a zero-mean Laplacian residual with scale parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _{\rho }=0.06~\textrm{m}$$\end{document} is assumed, together with occasional outliers affecting 5% of the measurements, where an additional perturbation uniformly distributed in [-0.25,0.25] m is added. This models imperfect delay extraction and sporadic timing distortions. The RSS measurement is modeled as44\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{P}_{r,i} = P_{r,i}(\textbf{p}) + n_i^{(P)}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_i^{(P)}$$\end{document} follows a Laplacian distribution with scale \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _P = 0.08\,P_{r,i}^{\textrm{max}}$$\end{document}, truncated to ensure physical admissibility, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{r,i}^{\textrm{max}}$$\end{document} denotes a nominal maximum received optical power used for normalization. To emulate realistic optical perturbations, the RSS observations further include:an ambient-light perturbation term representing background illumination variability,a multiplicative gain fluctuation of up to 3%, acting as a surrogate for LED power mismatch and front-end gain uncertainty,occasional partial attenuation events on 10% of the links, corresponding to mild LOS obstruction, partial shadowing, or angular mismatch. an ambient-light perturbation term representing background illumination variability, an ambient-light perturbation term representing background illumination variability, a multiplicative gain fluctuation of up to 3%, acting as a surrogate for LED power mismatch and front-end gain uncertainty, a multiplicative gain fluctuation of up to 3%, acting as a surrogate for LED power mismatch and front-end gain uncertainty, occasional partial attenuation events on 10% of the links, corresponding to mild LOS obstruction, partial shadowing, or angular mismatch. occasional partial attenuation events on 10% of the links, corresponding to mild LOS obstruction, partial shadowing, or angular mismatch. This disturbance model should therefore be interpreted as a phenomenological surrogate at the positioning-observable level. It captures non-Gaussian RSS corruption after propagation and front-end processing, but it does not explicitly reproduce NLOS multipath reflection, severe hard-shadowing events, PD saturation/clipping, or injection-dependent MQW LED nonlinearities, efficiency-droop effects, and waveform-level electro-optic distortion. Incorporating ray-tracing-based optical propagation or a physics-based compact LED model such as6 would require additional environment and device characterization data, and is left for future work. For each test point, M=50 measurement snapshots are generated and averaged at the decision stage, which reduces snapshot-level variance while still reflecting the impact of heavy-tailed residuals and occasional perturbation events in the aggregated observations. Compared methods The proposed algorithm is compared against four reference estimators: Optical pseudorange WNLS: a weighted nonlinear least-squares trilateration method using only pseudorange measurements and ignoring bias estimation.RSS-only Lambertian WNLS: a deterministic power-based estimator using only optical RSS and the Lambertian channel model.Joint pseudorange–RSS (PSO + Huber): a robust fusion approach using the same two observables and the same PSO search configuration, but replacing the proposed Laplacian NLL formulation with Huber-type penalties.Oracle estimator: a lower-bound reference assuming perfect knowledge of the bias vector and disturbance-free channel parameters. Optical pseudorange WNLS: a weighted nonlinear least-squares trilateration method using only pseudorange measurements and ignoring bias estimation. Optical pseudorange WNLS: a weighted nonlinear least-squares trilateration method using only pseudorange measurements and ignoring bias estimation. RSS-only Lambertian WNLS: a deterministic power-based estimator using only optical RSS and the Lambertian channel model. RSS-only Lambertian WNLS: a deterministic power-based estimator using only optical RSS and the Lambertian channel model. Joint pseudorange–RSS (PSO + Huber): a robust fusion approach using the same two observables and the same PSO search configuration, but replacing the proposed Laplacian NLL formulation with Huber-type penalties. Joint pseudorange–RSS (PSO + Huber): a robust fusion approach using the same two observables and the same PSO search configuration, but replacing the proposed Laplacian NLL formulation with Huber-type penalties. Oracle estimator: a lower-bound reference assuming perfect knowledge of the bias vector and disturbance-free channel parameters. Oracle estimator: a lower-bound reference assuming perfect knowledge of the bias vector and disturbance-free channel parameters. The estimator studied here corresponds to the joint optical delay–RSS formulation introduced in the Methods section, where the receiver position and the bias vector are jointly obtained through PSO. These baselines represent the principal families of existing VLP approaches, namely range-based positioning, RSS-based positioning, and robust multi-observable fusion. This choice enables a fair comparison of both accuracy and computational burden under the same room geometry, anchor layout, and disturbance conditions. In particular, the Joint pseudorange–RSS (PSO + Huber) benchmark helps isolate the benefit of the proposed estimator formulation from the optimizer choice, since both joint methods use the same observables and the same PSO-based search strategy. The Oracle estimator is included as an operational lower-bound reference. Although it is not a formal CRLB, it plays a similar benchmarking role in the present non-Gaussian, bias-aware setting, where a classical unbiased Gaussian bound would be only partially representative. Published PSO/RSS-based VLP methods are discussed in the Introduction as related work, but their reported numerical results are not transplanted directly as baselines because they usually rely on different room dimensions, optical front ends, and calibration assumptions. PSO configuration The particle swarm optimizer uses a swarm size of S=40 particles and a maximum of 120 iterations. The inertia factor is set to ω =0.7, while the cognitive and social coefficients are chosen as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_1=c_2=1.5$$\end{document}. The search space is constrained by the room geometry:45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x \in [0,8], \qquad y \in [0,11], \end{aligned}$$\end{document}and the per-link bias search interval is defined as46\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} b_i \in [-0.25,\,0.25]\;\textrm{m}, \quad i=1,\dots ,4. \end{aligned}$$\end{document}Because the receiver height is fixed at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z=0.85\,\textrm{m}$$\end{document} in the reported experiments, the effective PSO search dimension in this section is 2+N=6 rather than the generic 3+N formulation introduced in the Methods section. To accelerate convergence, the swarm is initialized around the solution of the Optical pseudorange WNLS estimator, in the same spirit as the initialization strategy adopted in the reference paper. This reduces the computational burden and improves the probability of convergence toward the global minimum. The adopted configuration S=40 and K=120 should be interpreted as a pragmatic compromise between global exploration and computational cost. Under this setting, the worst-case budget is S× K=4800 objective-function evaluations per position estimate. Increasing either the swarm size or the iteration cap improves search diversity only at a linearly increasing computational cost, whereas aggressive reductions in S or K would generally make the search more sensitive to multimodality and local minima. From a convergence standpoint, the Optical pseudorange WNLS solution acts as a geometry-consistent warm start that narrows the effective search region and reduces sensitivity to purely random initialization. Since the absolute wall-clock runtime depends strongly on software implementation, programming language, and hardware parallelism, the present manuscript reports platform-independent evaluation counts and representative asymptotic complexity rather than hardware-specific execution times. Evaluation metrics Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_j,y_j)$$\end{document} denote the true receiver coordinates at the j-th test point and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\hat{x}_j,\hat{y}_j)$$\end{document} the corresponding estimate. The positioning error (PE) is defined as47\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textrm{PE}_j = \sqrt{(\hat{x}_j-x_j)^2 + (\hat{y}_j-y_j)^2}. \end{aligned}$$\end{document}The following performance metrics are reported:RMSE,95th percentile positioning error,standard deviation of positioning error,CDF of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PE}$$\end{document}. RMSE, RMSE, 95th percentile positioning error, 95th percentile positioning error, standard deviation of positioning error, standard deviation of positioning error, CDF of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PE}$$\end{document}. CDF of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{PE}$$\end{document}. Main localization results Figure 2 shows the cumulative distribution function of the positioning error for all considered estimators in the main LOS scenario. It can be observed that the proposed joint optical delay–RSS estimator consistently outperforms the standalone baselines and also improves over Joint pseudorange–RSS (PSO + Huber) across the full error range. As shown in Fig. 2, this estimator shifts the entire error distribution toward lower values, indicating not only improved average accuracy but also stronger robustness against large-error events.Fig. 2CDF of the positioning error for the considered VLC estimators in the main LOS scenario. The proposed joint optical delay–RSS estimator yields the most favorable overall performance among the practical estimators, especially in the medium- and high-error regions of the distribution. CDF of the positioning error for the considered VLC estimators in the main LOS scenario. The proposed joint optical delay–RSS estimator yields the most favorable overall performance among the practical estimators, especially in the medium- and high-error regions of the distribution. The Optical pseudorange WNLS estimator exhibits the poorest performance among the non-Oracle estimators, mainly because the per-link optical timing biases are not explicitly estimated and are therefore absorbed as geometric inconsistency. The RSS-only Lambertian WNLS estimator performs better than Optical pseudorange WNLS in central regions of the room, but its performance degrades near corners and in geometries with reduced angular diversity, where multiple positions can produce similar received power patterns. The Joint pseudorange–RSS (PSO + Huber) benchmark improves robustness relative to both single-observable baselines, but still remains more sensitive than the proposed estimator to asymmetric and heavy-tailed measurement deviations. By jointly estimating the user position and the per-link bias vector under a Laplacian NLL framework, the proposed estimator achieves the most favorable CDF over the full error range. Table 1 summarizes the numerical performance. The proposed estimator attains an RMSE of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.24~\textrm{m}$$\end{document}, a 95th percentile error of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.46~\textrm{m}$$\end{document}, and a standard deviation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.13~\textrm{m}$$\end{document}. Compared with Optical pseudorange WNLS, this corresponds to an RMSE reduction of approximately 63.6% and a 95th percentile improvement of approximately 58.9%.Table 1Localization Performance in the Main LOS VLC Scenario.MethodRMSE (m)95th Perc. (m)Std. Dev. (m)Oracle estimator0.090.190.05Optical pseudorange WNLS0.661.120.31RSS-only Lambertian WNLS0.430.810.24Joint pseudorange–RSS (PSO + Huber)0.310.580.17Proposed joint optical delay–RSS estimator0.240.460.13 Table 1Localization Performance in the Main LOS VLC Scenario.MethodRMSE (m)95th Perc. (m)Std. Dev. (m)Oracle estimator0.090.190.05Optical pseudorange WNLS0.661.120.31RSS-only Lambertian WNLS0.430.810.24Joint pseudorange–RSS (PSO + Huber)0.310.580.17Proposed joint optical delay–RSS estimator0.240.460.13 Localization Performance in the Main LOS VLC Scenario. To complement the accuracy results, Table 2 summarizes the representative computational burden of the practical estimators. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{\textrm{W}}$$\end{document} denotes the number of iterations required by a local WNLS solver, whereas the PSO-based estimators require S× K particle evaluations. Under the adopted configuration, S=40 and K=120, so the PSO-based methods perform up to 4800 particle evaluations per position estimate. Accordingly, the proposed estimator is more computationally demanding than the single-observable WNLS baselines, but it offers the best localization accuracy while preserving the same asymptotic complexity order as Joint pseudorange–RSS (PSO + Huber).Table 2Accuracy and Representative Computational Burden of the Practical Estimators.MethodRMSE (m)OptimizationRepresentative ComplexityBurdenOptical pseudorange WNLS0.66Local WNLS\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(I_{\textrm{W}}N)$$\end{document}LowRSS-only Lambertian WNLS0.43Local WNLS\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(I_{\textrm{W}}N)$$\end{document}Low–moderateJoint pseudorange–RSS (PSO + Huber)0.31PSO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(SKN)$$\end{document}HighProposed joint optical delay–RSS estimator0.24PSO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(SKN)$$\end{document}High Table 2Accuracy and Representative Computational Burden of the Practical Estimators.MethodRMSE (m)OptimizationRepresentative ComplexityBurdenOptical pseudorange WNLS0.66Local WNLS\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(I_{\textrm{W}}N)$$\end{document}LowRSS-only Lambertian WNLS0.43Local WNLS\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(I_{\textrm{W}}N)$$\end{document}Low–moderateJoint pseudorange–RSS (PSO + Huber)0.31PSO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(SKN)$$\end{document}HighProposed joint optical delay–RSS estimator0.24PSO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {O}(SKN)$$\end{document}High Accuracy and Representative Computational Burden of the Practical Estimators. Overall, the results support the main intuition behind the joint formulation. The pseudorange observable provides direct geometric information but is vulnerable to unresolved timing bias. The RSS observable carries complementary geometric information through the Lambertian channel gain, but is sensitive to angular mismatch and power uncertainty. Their joint exploitation reduces ambiguity and improves resilience against localized measurement corruption. Spatial distribution of positioning error Figure 3 illustrates the spatial distribution of the mean positioning error obtained with the proposed estimator. The results show that the lowest error levels are achieved in the central area of the room, where the receiver benefits from more balanced anchor geometry and more favorable incidence angles. By contrast, error levels slightly increase near the room edges and corners, where the angular diversity decreases and the optical channel becomes more sensitive to geometry-induced distortion.Fig. 3Spatial distribution of the mean positioning error of the proposed estimator over the 48 reference points. Lower errors are observed in the central region of the room, whereas slightly larger errors appear near boundaries and corners due to less favorable optical geometry. Spatial distribution of the mean positioning error of the proposed estimator over the 48 reference points. Lower errors are observed in the central region of the room, whereas slightly larger errors appear near boundaries and corners due to less favorable optical geometry. Even in these less favorable regions, the proposed estimator maintains sub-meter performance. This behavior is important because RSS-only estimators often exhibit pronounced edge degradation in optical positioning due to reduced angular diversity and stronger sensitivity to receiver orientation. Impact of receiver orientation perturbation In practical VLC systems, receiver orientation is rarely fixed. To assess robustness to moderate orientation uncertainty, an additional experiment was conducted where the receiver tilt angles were randomly perturbed within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document} in both pitch and roll. The estimator itself was not provided with the exact orientation values and therefore operated under nominal upward-facing assumptions. Figure 4 presents the cumulative distribution function of the positioning error when the receiver experiences random tilt perturbations in pitch and roll. Although all methods degrade under orientation mismatch, the proposed joint optical delay–RSS estimator still delivers the most favorable error distribution among the practical schemes. This behavior indicates that the fusion of pseudorange and RSS observables provides improved resilience against moderate orientation uncertainty compared with single-observable VLC positioning methods.Fig. 4CDF of the positioning error under random receiver tilt perturbations of up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document} in pitch and roll. The proposed estimator yields the most favorable overall performance among the practical estimators despite the orientation mismatch. CDF of the positioning error under random receiver tilt perturbations of up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document} in pitch and roll. The proposed estimator yields the most favorable overall performance among the practical estimators despite the orientation mismatch. The corresponding numerical results are shown in Table 3. As expected, all methods degrade under orientation mismatch because the incidence angle directly affects the received optical power. However, the proposed joint optical delay–RSS estimator remains the best performing non-Oracle estimator. Its degradation is more graceful than that of RSS-only Lambertian WNLS because the optical pseudorange still provides geometry-related information even when the power pattern is distorted.Table 3Localization Performance Under Random Receiver Tilt (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document}).MethodRMSE (m)95th Perc. (m)Std. Dev. (m)Oracle estimator0.130.240.08Optical pseudorange WNLS0.711.240.35RSS-only Lambertian WNLS0.611.030.33Joint pseudorange–RSS (PSO + Huber)0.400.730.22Proposed joint optical delay–RSS estimator0.320.610.18 Table 3Localization Performance Under Random Receiver Tilt (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document}).MethodRMSE (m)95th Perc. (m)Std. Dev. (m)Oracle estimator0.130.240.08Optical pseudorange WNLS0.711.240.35RSS-only Lambertian WNLS0.611.030.33Joint pseudorange–RSS (PSO + Huber)0.400.730.22Proposed joint optical delay–RSS estimator0.320.610.18 Localization Performance Under Random Receiver Tilt (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document}). This observation is important from a practical perspective. It shows that even without explicitly estimating orientation, the proposed joint optical delay–RSS estimator provides a degree of immunity against moderate angular mismatch. This supports the relevance of combining timing and power observables rather than relying on power measurements alone. Perturbation coverage and sensitivity scope A fair concern is that the evaluated environment remains LOS-dominant and does not constitute a full ray-traced or hardware-calibrated VLC propagation study. However, the present numerical framework does probe several practically relevant impairments at the observation level. Table 4 summarizes how these effects are represented and clarifies which ones remain outside the current scope.Table 4Practical Perturbations Represented in the Current Evaluation.Practical effectRepresentation in this workWhere assessedAmbient-light variabilityAdditive ambient-light perturbation in RSS observationsMain LOS scenarioLED power mismatch/gain uncertaintyMultiplicative RSS fluctuation up to 3%Main LOS scenarioMild shadowing/partial blockagePartial attenuation events on 10% of linksMain LOS scenarioReceiver orientation mismatchRandom pitch/roll perturbations up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document}Table 3, Fig. 4Timing hardware offsetsPer-link bias model plus sweep from [\!-0.02,0.02] m to [\!-0.20,0.20] mFig. 5Not explicitly modeledNLOS ray-traced reflections, severe blockage, PD saturation/clipping, and physics-based LED electro-optic nonlinearityDiscussion/Future Work Table 4Practical Perturbations Represented in the Current Evaluation.Practical effectRepresentation in this workWhere assessedAmbient-light variabilityAdditive ambient-light perturbation in RSS observationsMain LOS scenarioLED power mismatch/gain uncertaintyMultiplicative RSS fluctuation up to 3%Main LOS scenarioMild shadowing/partial blockagePartial attenuation events on 10% of linksMain LOS scenarioReceiver orientation mismatchRandom pitch/roll perturbations up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 10^\circ$$\end{document}Table 3, Fig. 4Timing hardware offsetsPer-link bias model plus sweep from [\!-0.02,0.02] m to [\!-0.20,0.20] mFig. 5Not explicitly modeledNLOS ray-traced reflections, severe blockage, PD saturation/clipping, and physics-based LED electro-optic nonlinearityDiscussion/Future Work Practical Perturbations Represented in the Current Evaluation. Taken together, the nominal perturbation model, the receiver-tilt experiment, and the timing-bias sweep should therefore be interpreted as a structured sensitivity analysis along three main axes: amplitude-domain corruption, orientation mismatch, and synchronization-related bias. These tests strengthen the robustness discussion, but they should not be interpreted as a substitute for full NLOS-aware optical channel modeling or saturation-aware receiver characterization. Additional comparative results To further highlight the advantages of the proposed estimator, Table 5 reports the relative improvement of the proposed estimator with respect to the three practical baselines in the nominal LOS scenario and under receiver tilt. The entries in this table are not generated from additional simulations; instead, they are obtained by post-processing the values already reported in Tables 1 and 3. Specifically, for a metric \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta \in \{\textrm{RMSE},\textrm{P95}\}$$\end{document} and a given baseline, the percentage gain is computed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$G_{\eta }=100\times (\eta _{\textrm{base}}-\eta _{\textrm{prop}})/\eta _{\textrm{base}}$$\end{document}. For example, in the main LOS scenario, the RMSE gain with respect to the optical pseudorange WNLS baseline is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100\times (0.66-0.24)/0.66=63.6\%$$\end{document}. This normalization makes the advantage of the proposed formulation easier to interpret than from raw error values alone.Table 5Relative Improvement of the Proposed Estimator over the Practical Baselines.BaselineMain RMSE Gain (%)Main 95th Gain (%)Tilt RMSE Gain (%)Tilt 95th Gain (%)Optical pseudorange WNLS63.658.954.950.8RSS-only Lambertian WNLS44.243.247.540.8Joint pseudorange–RSS (PSO + Huber)22.620.720.016.4 Table 5Relative Improvement of the Proposed Estimator over the Practical Baselines.BaselineMain RMSE Gain (%)Main 95th Gain (%)Tilt RMSE Gain (%)Tilt 95th Gain (%)Optical pseudorange WNLS63.658.954.950.8RSS-only Lambertian WNLS44.243.247.540.8Joint pseudorange–RSS (PSO + Huber)22.620.720.016.4 Relative Improvement of the Proposed Estimator over the Practical Baselines. Several observations can be drawn from Table 5. First, the gains are consistently strongest with respect to the single-observable baselines. In the main LOS case, the proposed estimator improves the RMSE by 63.6% over Optical pseudorange WNLS and by 44.2% over RSS-only Lambertian WNLS, while the corresponding gains in 95th-percentile error are 58.9% and 43.2%, respectively. Under receiver tilt, the same trend remains visible, with RMSE gains of 54.9% over Optical pseudorange WNLS and 47.5% over RSS-only Lambertian WNLS. This indicates that the benefit of the joint estimator does not stem only from improved nominal fitting, but also from stronger robustness when the geometry or power pattern becomes less favorable. Second, the gains over the Joint pseudorange–RSS (PSO + Huber) baseline are more moderate but still systematic: 22.6% and 20.7% in RMSE and 95th-percentile error in the main scenario, and 20.0% and 16.4% under tilt. This is an important result because both methods rely on the same pair of observables and the same PSO search principle, so the observed improvement can be attributed primarily to the proposed bias-aware Laplacian formulation rather than to the use of a different optimization engine. In addition, Table 6 provides an aggregate cross-scenario view by combining the main and tilt experiments. The aggregate values are again obtained by post-processing previously reported results: the average RMSE, average 95th-percentile error, and average standard deviation are computed as simple arithmetic means over the main and tilt scenarios, whereas the worst-case RMSE is defined as the maximum of the two scenario-specific RMSE values. Thus, if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta _{\textrm{main}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta _{\textrm{tilt}}$$\end{document} denote the corresponding metrics, then \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\eta }=(\eta _{\textrm{main}}+\eta _{\textrm{tilt}})/2$$\end{document}, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta _{\textrm{worst}}=\max \{\eta _{\textrm{main}},\eta _{\textrm{tilt}}\}$$\end{document}. The proposed estimator attains the lowest average RMSE, the lowest worst-case RMSE, and the lowest average 95th-percentile error among the practical estimators, which confirms that its advantage is not restricted to a single test condition.Table 6Aggregate Cross-Scenario Performance Summary.MethodAvg. RMSE (m)Worst-Case RMSE (m)Avg. 95th Perc. (m)Avg. Std. Dev. (m)Optical pseudorange WNLS0.690.711.180.33RSS-only Lambertian WNLS0.520.610.920.29Joint pseudorange–RSS (PSO + Huber)0.360.400.660.20Proposed joint optical delay–RSS estimator0.280.320.540.16 Table 6Aggregate Cross-Scenario Performance Summary.MethodAvg. RMSE (m)Worst-Case RMSE (m)Avg. 95th Perc. (m)Avg. Std. Dev. (m)Optical pseudorange WNLS0.690.711.180.33RSS-only Lambertian WNLS0.520.610.920.29Joint pseudorange–RSS (PSO + Huber)0.360.400.660.20Proposed joint optical delay–RSS estimator0.280.320.540.16 Aggregate Cross-Scenario Performance Summary. Table 6 also provides a more global interpretation of the results. The proposed estimator yields an average RMSE of 0.28 m, compared with 0.36 m for Joint pseudorange–RSS (PSO + Huber), 0.52 m for RSS-only Lambertian WNLS, and 0.69 m for Optical pseudorange WNLS. These correspond to average RMSE reductions of 22.2%, 46.2%, and 59.4%, respectively. Because the proposed estimator and Joint pseudorange–RSS (PSO + Huber) share the same observables and the same PSO search strategy, the improvement over that baseline indicates that the gain cannot be attributed to PSO alone, but rather to the proposed likelihood and bias-aware modeling choices. A similar ranking is observed in the average 95th-percentile error, where the proposed estimator reaches 0.54 m, versus 0.66 m, 0.92 m, and 1.18 m for the same baselines. The worst-case RMSE further shows that the proposed estimator remains the best practical method even under the less favorable of the two evaluated conditions, with 0.32 m compared with 0.40 m for Joint pseudorange–RSS (PSO + Huber), 0.61 m for RSS-only Lambertian WNLS, and 0.71 m for Optical pseudorange WNLS. Finally, the lowest average standard deviation confirms that this estimator is not only more accurate on average, but also more stable. Taken together, these two additional summaries reinforce the main conclusion of the paper: the proposed robust joint optical delay–RSS estimator consistently yields the most favorable practical performance, both in terms of direct accuracy gains over established baselines and in terms of aggregate robustness across distinct operating conditions. Sensitivity to timing bias magnitude A central motivation of the proposed estimator is the explicit estimation of the per-link bias vector. To evaluate the importance of this feature, we varied the bias interval from [-0.02,0.02] m to [-0.20,0.20] m while keeping the remaining conditions fixed. Figure 5 reports the RMSE as a function of the maximum absolute timing bias. As the bias magnitude increases, Optical pseudorange WNLS deteriorates rapidly, whereas the proposed estimator remains significantly more stable thanks to the joint estimation of the per-link bias vector. This behavior directly validates the inclusion of the bias parameters in the optimization problem and confirms that bias-aware fusion is particularly beneficial in timing-assisted VLC localization.Fig. 5RMSE versus maximum absolute optical timing bias for the considered estimators. The proposed estimator exhibits a slower performance degradation than the reference methods due to its explicit bias-aware joint estimation framework. RMSE versus maximum absolute optical timing bias for the considered estimators. The proposed estimator exhibits a slower performance degradation than the reference methods due to its explicit bias-aware joint estimation framework. As the timing bias magnitude increases, Optical pseudorange WNLS deteriorates rapidly because the geometric model becomes increasingly inconsistent. In contrast, the proposed estimator exhibits a much slower performance degradation thanks to the joint estimation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{b}$$\end{document}. Joint pseudorange–RSS (PSO + Huber) also benefits from multimodal information, but still underperforms the proposed estimator because it does not model the residual structure as effectively under the heavy-tailed optical error conditions considered here. Discussion The results support three main conclusions. First, joint delay–RSS fusion is beneficial because the pseudorange observable contributes distance-like geometric constraints while RSS adds angle- and gain-dependent discrimination through the Lambertian channel law. Second, explicit bias estimation is important in timing-assisted VLC localization: the bias-sensitivity analysis shows that the advantage of the proposed estimator becomes more pronounced as the bias magnitude increases. Third, the Laplacian NLL formulation improves robustness to non-Gaussian optical disturbances, particularly in the upper tail of the error distribution. Overall, the obtained sub-meter accuracy is consistent with a structured LOS environment with moderate measurement corruption. At the same time, the results should be interpreted within the methodological scope of the paper: the scenario remains simulation-based and LOS-dominant, and the PSO-based solution trades computational cost for global-search robustness. From a complexity perspective, the proposed estimator is computationally heavier than the WNLS baselines, but it does not incur a higher asymptotic order than Joint pseudorange–RSS (PSO + Huber) and achieves better localization accuracy under the same optimization framework. Despite these encouraging results, several limitations should be acknowledged. First, the evaluation is based on a simulated LOS-dominant environment with four anchors and controlled perturbations, which may not fully capture the variability of real deployments. Second, receiver orientation is treated through perturbation tests rather than explicit orientation estimation, leaving room for further improvement in handheld scenarios. Third, although PSO improves robustness in nonconvex settings, its computational cost may increase with larger search spaces and denser anchor deployments. In addition, experimental verification of the synchronization-assisted optical delay component would require dedicated high-resolution timing instrumentation that is beyond the scope of the present study. We also note that there is currently no widely adopted public dataset, to the best of our knowledge, that jointly provides synchronized VLC pseudorange-like delay observations, RSS measurements, and per-link timing-bias effects under a known optical anchor geometry. The present RSS model is also phenomenological with respect to LED device physics: while it captures aggregate distortion at the measurement level, it does not explicitly include the injection-dependent electro-optic nonlinearity of GaN MQW LEDs reported in recent device-level VLC studies6. Finally, the observability arguments of the present estimator rely on a fixed-height 2D configuration with quasi-static per-link biases; extending the same formulation to fully 3D operation, unknown orientation, or rapidly drifting biases would require additional constraints such as more anchors or temporal priors. The current PSO implementation should therefore be interpreted as an algorithmic validation tool rather than as a finalized real-time embedded solution. Experimental validation in real VLC testbeds and complexity-aware implementations therefore remain important next steps. Conclusion This paper presented a robust joint optical delay–RSS positioning framework for indoor VLC systems. This approach was developed to address practical VLC localization challenges arising from timing bias, ambient-light perturbations, receiver orientation sensitivity, and aggregate front-end/RSS distortions. To this end, synchronization-assisted optical delay pseudoranges and Lambertian RSS measurements were fused through a weighted negative log-likelihood formulation with explicit per-link bias estimation. Because the resulting estimation problem is nonlinear and multimodal, particle swarm optimization was adopted to jointly estimate the receiver position and the bias vector without relying on gradient information. This design allows the method to remain effective under non-Gaussian optical measurement disturbances and realistic model mismatch conditions. The reported numerical results showed that the proposed estimator outperforms representative pseudorange-only, RSS-only, and alternative robust fusion baselines in the considered VLC indoor scenario. In addition to improving the overall error distribution, the method also exhibited stronger robustness under receiver tilt perturbations and increased timing bias. Within the considered LOS-dominant, simulation-based setting, these results confirm that the joint exploitation of optical delay and RSS observables is an effective strategy for improving visible light positioning robustness and accuracy. Within the adopted modeling assumptions, this framework therefore provides a robust and technically grounded basis for indoor VLC positioning and a promising starting point for future experimental validation in smart and industrial optical wireless environments. Future work Future work will focus on extending the proposed framework in several directions. First, explicit receiver orientation estimation should be incorporated in order to better capture practical handheld and mobile-device operation. Second, more advanced optical propagation models including NLOS components and partial blockage effects should be considered. Third, physics-based compact models of LED electro-optic nonlinearity should be integrated into the RSS-generation stage so that injection-dependent device behavior can be propagated to the positioning observables more faithfully. Fourth, the proposed method should be validated experimentally using real VLC testbeds, synchronized optical beacons, and dedicated high-resolution timing hardware for optical delay extraction. Fifth, implementation-oriented acceleration strategies such as parallel particle evaluation, adaptive swarm sizing, and early-stopping criteria should be investigated for real-time optical positioning. Finally, adaptive or learning-assisted weighting and residual modeling strategies may further improve localization robustness in dynamic indoor environments.