betaFeedback

WorksComputer Science

A Tightly Coupled Magnetic-Inertial SLAM System for Field Robots

Abstract

Field robots operating in mines, tunnels, and disaster areas require reliable localization despite GNSS denial and degraded camera and LiDAR measurements caused by dust, smoke, or poor illumination. Missions involving both ground and aerial robots further motivate complementary sensing modalities that do not depend on visibility, preinstalled infrastructure, or wheel contact. This letter presents MISLAM, a tightly coupled magnetic-inertial simultaneous localization and mapping system using an IMU and a nine-magnetometer array, without relying on external odometry or platform-specific motion assumptions. The system combines the inertial navigation model with local polynomial magnetic-field models for short-term motion estimation and a global Gaussian process model for drift correction using an error-state Kalman filter. Offline evaluation on three indoor datasets collected with a handheld rig shows reductions in total endpoint closure error of 45.7% - 78.0% relative to magnetic-inertial odometry. Under poor illumination, MISLAM maintains tracking after visual-inertial odometry fails, achieving a horizontal closure error of 0.332 m, although vertical drift remains a limitation. These results support magnetic-inertial SLAM as a complementary localization modality for robotic sensor suites operating under degraded visual conditions.

Introduction

Reliable localization is a prerequisite for field robots exploring mines, tunnels, and disaster areas. These environments often deny access to global navigation satellite systems (GNSS), while poor illumination, dust, and smoke can degrade camera- and LiDAR-based localization [1]. These challenges motivate complementary sensing modalities that provide motion information without relying on visibility, preinstalled infrastructure, or wheel contact. Such capabilities are relevant to both ground and aerial robots and can support localization when other sensors become unreliable.

Indoor magnetic fields have been recognized as a promising environmental feature for localization because of their spatial inhomogeneity and temporal stability. Ferromagnetic structures and objects create spatial variations that can remain sufficiently stable to serve as location-dependent signatures [10]. Furthermore, magnetic-field-based localization does not require capturing identifiable visual information, avoiding the associated image-privacy concerns. These properties have motivated sustained research over the past two decades, spanning magnetic-field fingerprinting and model-based localization methods [3, 4, 6, 7, 8, 13, 15, 16, 22, 23, 24, 25, 26, 27].

Magnetic fingerprinting typically matches measurements to locations within a pre-surveyed map. However, surveying an unknown environment in advance can be time-consuming and is often infeasible prior to a field robotics mission. Magnetic-field SLAM addresses this issue by estimating the trajectory and field map jointly. Representative methods [6, 22, 24] use Gaussian processes (GPs) to model the magnetic field along the trajectory and constrain the drift accumulated during pose propagation. Despite these advances, existing magnetic SLAM methods still rely on accurate external odometry—typically from visual-inertial odometry (VIO) or pedestrian dead reckoning—which limits their applicability when primary odometry sensors degrade or the platform’s motion pattern is incompatible with the assumptions of the odometry method. Achieving self-contained magnetic-field SLAM without external odometry or restrictive motion assumptions presents two main challenges: 1) High computational complexity: Extracting odometry directly from magnetic field measurements requires high-resolution field models capable of capturing sub-meter spatial variations. As reported in [19], 50 to 500 basis functions per square meter are needed to support accurate localization. 2) Degenerate spatial sampling: Spatially proximate measurements often yield overconfident predictions of the magnetic field at nearby locations, particularly when entering unexplored areas [9, 23]. Mitigating this issue requires enforcing a minimum guard distance or time interval between map updates, which further complicates high-resolution modeling.

Magnetometer arrays enable an alternative, local modeling of the magnetic fields across the array. Because the array covers a relatively small volume, accurate yet computationally efficient local field models can be constructed. Moreover, because the relative positions of the magnetometers are fixed and known, the local field model can be fitted and evaluated from a single multisensor snapshot with lower uncertainty than fitting single-magnetometer measurements along an estimated trajectory. This type of modeling enables magnetic odometry inspired by optical flow [18]. Magnetic-inertial odometry (MIO) systems, including MAINS [3] and other array-based methods [11, 12, 25, 28], fuse array measurements with inertial measurements to bound velocity and position drift. Although these methods are self-contained and require neither external odometry nor motion constraints, existing MIO systems do not maintain a persistent global field map. Consequently, they cannot exploit loop closures when revisiting mapped areas, leaving long-term navigation drift uncorrected.

Building on these developments, this letter presents MISLAM, a tightly coupled magnetic-inertial SLAM system that combines the complementary strengths of local polynomial field modeling and global GP mapping using a single error-state Kalman filter (ESKF). The main contributions of this letter are as follows:

  • We develop a self-contained magnetic-inertial SLAM system using an IMU and a nine-magnetometer array that operates without external odometry or restrictive motion constraints.

  • We develop an ESKF update strategy that validates local field references using chi-square tests, retires each reference after one update attempt, and selects global GP keyframes based on the number of accepted local updates.

  • We evaluate MISLAM against a state-of-the-art MIO system (MAINS [3]) and a VIO system (OpenVINS [2]) across three indoor datasets collected with a handheld sensor rig under varying revisit topologies and lighting conditions.

Overview

The proposed MISLAM system fuses inertial and magnetic measurements to jointly estimate the trajectory and the magnetic-field map. The system uses a sensor rig that consists of an array of NN magnetometers and an IMU. The processing stages—sensor calibration, stationary initialization, inertial propagation, and local and global magnetic updates—are shown in Figure 1 and detailed in Sections 3 and 4. Here, we present the main concepts and notation.

Figure 1. System overview of the tightly coupled magnetic-inertial SLAM architecture.

After proper sensor calibration and filter initialization, IMU measurements propagate the position, velocity, and attitude. Magnetic measurements from two disjoint groups—the local group LL and the global group GG—reduce the propagation error through the local and global field representations, respectively. A local model captures the spatial field variation around a recent reference pose and provides short-term relative motion information. A global model retains magnetic information across the explored environment, allowing measurements in previously mapped regions to reduce accumulated navigation drift.

Throughout this letter, WW denotes the navigation world frame, WcalW_{\mathrm{cal}} denotes the calibration world frame, bb denotes the body frame coinciding with the IMU frame, and MiM_i denotes the coordinate frame of the iith magnetometer. The world-frame origin is placed at the initial IMU position. Its zz-axis points upward, opposite to gravity, such that gW=[0,0,−g]T\mathbf{g}^{W} = [0, 0, -g]^{\mathsf{T}}. The frame WcalW_{\mathrm{cal}} has its zz-axis pointing upward and its yy-axis aligned with the horizontal component of the calibration magnetic field. A rotation matrix RABR^{B}_{A} maps vectors from frame AA to frame BB; in particular, Rk=RbkWR_k = R^{W}_{b_k} denotes the body-to-world rotation at time kk. The fixed displacement from the IMU origin to magnetometer ii, expressed in the body frame, is denoted by di\mathbf{d}_i.

System Initialization

Magnetometer-IMU Joint Calibration

Accurate magnetic-inertial fusion requires consistent magnetic measurements expressed relative to the IMU frame. Magnetometer measurements are affected by hard-iron offsets, soft-iron distortions, scale-factor errors, and nonorthogonal sensing axes. Mounting errors also introduce misalignment between each magnetometer and the IMU. We account for these effects using the joint calibration method in [5].

Calibration data are collected by slowly rotating the assembled sensor rig through a wide range of orientations in an open outdoor area, where the magnetic field is approximately constant. Translational acceleration is kept small so that the accelerometer measurements are dominated by gravity. Within this subsection, let Rk≜RbkWcalR_k \triangleq R^{W_{\mathrm{cal}}}_{b_k} denote the rotation from the IMU frame to the calibration world frame. The orientation dynamics and measurements of the sensor rig are modeled as

Rk+1=RkExp⁡([(ω~k−oω−ekω)Δt]×),(1)R_{k+1} = R_k \operatorname{Exp}\left([\left(\tilde{\boldsymbol{\omega}}_k - \mathbf{o}^{\omega} - \mathbf{e}^{\omega}_k\right)\Delta t]_{\times}\right), \tag*{(1)}
a~k=−RkTgWcal+oa+eka,(2)\tilde{\mathbf{a}}_k = -R_k^{\mathsf{T}}\mathbf{g}^{W_{\mathrm{cal}}} + \mathbf{o}^{a} + \mathbf{e}^{a}_k, \tag*{(2)}
m~ki=CiRkTBcalWcal(ι)+omi+ekmi,i=1,…,N,(3)\tilde{\mathbf{m}}^i_k = C_i R_k^{\mathsf{T}}\mathbf{B}^{W_{\mathrm{cal}}}_{\mathrm{cal}}(\iota) + \mathbf{o}^{m_i} + \mathbf{e}^{m_i}_k,\quad i = 1,\ldots,N, \tag*{(3)}

where [⋅]×[\cdot]_{\times} denotes the operator that maps R3\mathbb{R}^{3} to the Lie algebra so(3)\mathfrak{so}(3). The variables ω~k\tilde{\boldsymbol{\omega}}_k and a~k\tilde{\mathbf{a}}_k denote the gyroscope and accelerometer measurements, respectively, with corresponding biases oω\mathbf{o}^{\omega} and oa\mathbf{o}^{a}. The sampling interval is denoted by Δt\Delta t. For the iith magnetometer, m~ki\tilde{\mathbf{m}}^i_k and omi\mathbf{o}^{m_i} denote its measurement and bias, respectively. The vector BcalWcal(ι)=[0 cos⁡(ι) −sin⁡(ι)]T\mathbf{B}^{W_{\mathrm{cal}}}_{\mathrm{cal}}(\iota) = [0\ \cos(\iota)\ -\sin(\iota)]^{\mathsf{T}} denotes the normalized magnetic field expressed in the calibration world frame, where ι\iota is the dip angle. The fixed matrix Ci=DiRbMiC_i = D^i R^{M_i}_{b} combines the intrinsic distortion matrix DiD^i and the rotation RbMiR^{M_i}_{b} from the IMU frame to the iith magnetometer frame. For all sensors, the terms ek(⋅)\mathbf{e}^{(\cdot)}_k denote measurement noise, which is assumed to be white, zero-mean Gaussian with known covariance Σ(⋅)\Sigma_{(\cdot)}.

Let θ≜(oa,oω,{Ci,omi}i,ι)\theta\triangleq(o^{a}, o^{\omega}, \{C_i, o^{mi}\}_i, \iota) denote the calibration parameters and R0:TR_{0:T} the orientation trajectory. The calibration parameters and orientation trajectory are jointly estimated by minimizing the following maximum-likelihood objective:

(θ^,R^0:T)ML≜arg⁡min⁡θ,R0:T L(θ,R0:T)(4)(\hat{\theta}, \hat{R}_{0:T})^{\mathrm{ML}} \triangleq\underset{\theta,R_{0:T}}{\arg\min}\ L(\theta,R_{0:T}) \tag*{(4)}

where

L(θ,R0:T)=∑k∥a~k+RkTgcalW−oa∥Σa2+∑k∥ω~k−Log⁡(RkTRk+1)Δt−oω∥Σω2+∑i∑k∥m~ki−CiRkTBcalW(ι)−omi∥Σmi2.(5)\begin{aligned} L(\theta,R_{0:T}) ={}& \sum_k \left\lVert\tilde{a}_k + R_k^{\mathsf{T}}g_{\mathrm{cal}}^{W} - o^a \right\rVert_{\Sigma_a}^{2} \\ &+ \sum_k \left\lVert\tilde{\omega}_k - \frac{\operatorname{Log}(R_k^{\mathsf{T}}R_{k+1})}{\Delta t} - o^\omega\right\rVert_{\Sigma_\omega}^{2} \\ &+ \sum_i \sum_k \left\lVert\tilde{m}_k^i - C_iR_k^{\mathsf{T}}B_{\mathrm{cal}}^{W}(\iota) - o^{mi} \right\rVert_{\Sigma_{mi}}^{2}. \tag*{(5)} \end{aligned}

Here Log⁡:SO(3)→R3\operatorname{Log}: SO(3) \to\mathbb{R}^{3} denotes the rotation-vector logarithm and ∥r∥Σ2≜rTΣ−1r\lVert r\rVert_{\Sigma}^{2} \triangleq r^{\mathsf{T}}\Sigma^{-1}r. The estimated calibration parameters θ^\hat{\theta} are then used in the subsequent pose initialization and state estimation stages.

Stationary Pose Initialization

After joint magnetometer-IMU calibration, the rig is held stationary for initialization. The position and velocity are set to zero. The initial attitude is derived from accelerometer measurements. Let aˉ0\bar{a}_0 denote the mean bias-corrected accelerometer measurements during this phase. Under stationary conditions, aˉ0≈−R0TgW\bar{a}_0 \approx-R_0^{\mathsf{T}}g^{W}. We parameterize the initial attitude as R0=Rz(ψ0)Ry(ϑ0)Rx(ϕ0)R_0 = R_z(\psi_0)R_y(\vartheta_0)R_x(\phi_0). The initial roll and pitch are therefore [14]

ϕ0=atan2⁡(aˉ0,y,aˉ0,z),ϑ0=atan2⁡(−aˉ0,x,aˉ0,y2+aˉ0,z2).(6)\begin{aligned} \phi_0 &= \operatorname{atan2}(\bar{a}_{0,y},\bar{a}_{0,z}), \\ \vartheta_0 &= \operatorname{atan2}\left(-\bar{a}_{0,x},\sqrt{\bar{a}_{0,y}^{2}+\bar{a}_{0,z}^{2}}\right). \tag*{(6)} \end{aligned}

By convention, we set ψ0=0\psi_0 = 0, so that the initial heading of the sensor rig defines the horizontal orientation of the world frame.

Magnetic-Inertial SLAM System

The proposed MISLAM system combines inertial propagation and observations of indoor magnetic fields, on both local and global scales, in a single state-space model and uses the error-state Kalman filter [20] for state estimation. After initialization, each incoming IMU sample propagates the navigation state and joint covariance of the navigation state, global magnetic field model coefficients, and the reference pose when available. When an eligible local magnetic reference is available, a local magnetic update is attempted. The estimator then evaluates the keyframe selection criteria. At selected keyframes, a global magnetic update is attempted. The online recursion is summarized as follows:

  1. Propagate the navigation state and joint covariance with the IMU.

  1. Create or validate a local field reference, or attempt an update using an eligible reference. Retire the reference after one update attempt, whether the update is accepted or rejected.

  1. Check the accepted local magnetic update count to determine whether a new keyframe is selected.

  1. At the selected keyframe, attempt the global magnetic update.

  1. Output the current navigation estimate and prepare the joint state for the next sample.

The following subsections describe the state estimation algorithm. Unless otherwise indicated, state variables denote nominal estimates. Superscripts minus and plus denote the prior and posterior for the current update.

Joint State and Error State

The persistent state xkpx_k^p consists of the navigation state xn,kx_{n,k} and the global magnetic-field model coefficients uku_k:

xkp≜(xn,k,uk),xn,k≜(pk,vk,Rk,oka,okω).(7)\begin{aligned} x_k^p &\triangleq(x_{n,k},u_k),\\ x_{n,k} &\triangleq(p_k,v_k,R_k,o_k^a,o_k^\omega). \tag*{(7)} \end{aligned}

Here pkp_k and vkv_k are the IMU position and velocity in the world frame. The vector uku_k contains the magnetic field model coefficients defined in (15). The corresponding error state is

δxkp≜[δxn,k⊤,δuk⊤]⊤,δxn,k≜[δpk⊤,δvk⊤,δϕk⊤,δoka⊤,δokω⊤]⊤.(8)\begin{aligned} \delta x_k^p &\triangleq[\delta x_{n,k}^{\top},\delta u_k^{\top}]^{\top},\\ \delta x_{n,k} &\triangleq[\delta p_k^{\top},\delta v_k^{\top},\delta\phi_k^{\top},\delta o_k^{a\top},\delta o_k^{\omega\top}]^{\top}. \tag*{(8)} \end{aligned}

Vector errors are additive, while the three-dimensional attitude error is defined by the world-frame left perturbation Rktrue=Exp⁡([δϕk]×)RkR_k^{\mathrm{true}}=\operatorname{Exp}([\delta\phi_k]_{\times})R_k.

To estimate relative motion from measurements at different epochs, the filter augments the persistent state with a temporary local-reference pose ξr=(pr,Rr)\xi_r=(p_r,R_r), which is a clone of a past navigation pose and contributes six error coordinates, δξr=[δpr⊤,δϕr⊤]⊤\delta\xi_r=[\delta p_r^{\top},\delta\phi_r^{\top}]^{\top}, with the same attitude-error convention. When the temporary pose exists, the augmented state and its error state are

xk=(xkp,ξr),δxk=[(δxkp)⊤,δξr⊤]⊤.(9)\begin{aligned} x_k &= (x_k^p,\xi_r),\\ \delta x_k &= [(\delta x_k^p)^{\top},\delta\xi_r^{\top}]^{\top}. \tag*{(9)} \end{aligned}

The local-reference block is omitted when no local reference is active. The joint covariance Pk=cov⁡(δxk)P_k=\operatorname{cov}(\delta x_k) includes all cross-covariances between the navigation errors, coefficient errors, and local-reference pose errors. Pose augmentation and removal are described in § 4.5.

The local magnetic field model coefficients are fitted separately and are not included in the joint state.

Inertial Propagation

Define the bias-corrected IMU measurements fk=a~k−okaf_k = \tilde{a}_k - o^a_k and ωk=ω~k−okω\omega_k = \tilde{\omega}_k - o^\omega_k. Over a sampling interval Δt\Delta t, the nominal navigation state is propagated as

pk+1−=pk+vkΔt+12(Rkfk+gW)Δt2,vk+1−=vk+(Rkfk+gW)Δt,Rk+1−=RkExp⁡([ωkΔt]×).(10)\begin{aligned} p_{k+1}^{-} &= p_k + v_k\Delta t + \frac{1}{2}(R_k f_k + g^W)\Delta t^2,\\ v_{k+1}^{-} &= v_k + (R_k f_k + g^W)\Delta t,\\ R_{k+1}^{-} &= R_k\operatorname{Exp}([\omega_k\Delta t]_\times). \tag*{(10)} \end{aligned}

The IMU biases, okao^a_k and okωo^\omega_k, are modeled by two independent random walk processes. Linearizing the inertial error dynamics and applying Van Loan discretization yields the navigation transition Fn,kF_{n,k} and process covariance Qn,kQ_{n,k}; for details see [20 Ch. 7]. Map coefficients and pose clones have identity dynamics and no process noise. Thus, with identity and zero blocks sized to the active static states,

Fk=blkdiag⁡(Fn,k,I),F_k = \operatorname{blkdiag}(F_{n,k}, I),
Pk+1−=FkPk+FkT+blkdiag⁡(Qn,k,0).(11)P_{k+1}^{-} = F_k P_k^{+} F_k^{\mathsf{T}} + \operatorname{blkdiag}(Q_{n,k}, 0). \tag*{(11)}

Magnetic Models and Updates

Local Magnetic Model and Update

Within a small region around a reference pose, the local magnetic field is represented by the polynomial model [17], which is derived from a polynomial scalar potential and constrained to satisfy the divergence-free condition. For a polynomial model of degree nn, the resulting linear representation has Np=n2+4n+3N_p = n^2 + 4n + 3 coefficients and is of the following form:

B(s)=Φ(s)ηr,Φ(s)∈R3×Np,ηr∈RNp,(12)B(s) = \Phi(s)\eta_r,\qquad\Phi(s) \in\mathbb{R}^{3 \times N_p},\qquad\eta_r \in\mathbb{R}^{N_p}, \tag*{(12)}

where ss and B(s)B(s) denote the position and magnetic field vector, both expressed in the body frame at reference epoch rr, and ηr\eta_r denotes the local model coefficients.

Stacking the measurements of the local group LL into yry_r gives the linear model yr=Aηr+om+ery_r = A\eta_r + o^m + e_r, with blocks Ai=CiΦ(di)A_i = C_i\Phi(d_i), oim=oimo_i^m = o_i^m for i∈Li \in L. The vector ere_r includes the sensor and modeling errors. A weighted least-squares fit yields

η^r=PηrATΣfit−1(yr−om),(13)\hat{\eta}_r = P_{\eta_r} A^{\mathsf{T}}\Sigma_{\mathrm{fit}}^{-1}(y_r - o^m), \tag*{(13)}

where Pηr≜cov⁡(η^r)=(ATΣfit−1A)−1P_{\eta_r} \triangleq\operatorname{cov}(\hat{\eta}_r) = (A^{\mathsf{T}}\Sigma_{\mathrm{fit}}^{-1}A)^{-1} and Σfit≜cov⁡(er)\Sigma_{\mathrm{fit}} \triangleq\operatorname{cov}(e_r).

A reference is accepted only if its normalized fit residual passes a chi-square test,

rfitTΣfit−1rfit≤χ3NL−Np, 1−αfit2.(14)r_{\mathrm{fit}}^{\mathsf{T}}\Sigma_{\mathrm{fit}}^{-1}r_{\mathrm{fit}} \le\chi^2_{3N_L-N_p,\,1-\alpha_{\mathrm{fit}}}. \tag*{(14)}

Here rfit≜yr−Aη^r−omr_{\mathrm{fit}} \triangleq y_r - A\hat{\eta}_r - o^m denotes the fit residuals and NL≜∣L∣N_L \triangleq|L| denotes the number of magnetometers in LL. The accepted reference’s pose is then cloned into the joint state.

At a later epoch kk, define the relative rotation and translation

Rrk≜Rbkbr,trk=RrT(pk−pr).(15)R_{rk} \triangleq R_{bk}^{br}, \qquad t_{rk} = R_r^{\mathsf{T}}(p_k-p_r). \tag*{(15)}

The position of the iith magnetometer in the reference frame and its measurement model are

si,rk=trk+Rrkdi,s_{i,rk}=t_{rk}+R_{rk}d_i,
yi,k=hi,kloc(xk;ηr)+ei,kloc,(16)y_{i,k}=h_{i,k}^{\mathrm{loc}}(x_k;\eta_r)+e_{i,k}^{\mathrm{loc}}, \tag*{(16)}
hi,kloc(xk;ηr)≜CiRrkTΦ(si,rk)ηr+omi.h_{i,k}^{\mathrm{loc}}(x_k;\eta_r) \triangleq C_iR_{rk}^{\mathsf{T}}\Phi(s_{i,rk})\eta_r+o^{m_i}.

Here ei,kloce_{i,k}^{\mathrm{loc}} includes sensor noise and local model extrapolation errors. For the stacked measurements and an estimated η^r\hat{\eta}_r, the effective noise covariance is

Σloc,k=ΣL,kloc+Jη,kPηrJη,kT,(17)\Sigma_{\mathrm{loc},k}=\Sigma_{L,k}^{\mathrm{loc}}+J_{\eta,k}P_{\eta_r}J_{\eta,k}^{\mathsf{T}}, \tag*{(17)}

where

ΣL,kloc=cov⁡(eL,kloc),\Sigma_{L,k}^{\mathrm{loc}}=\operatorname{cov}(e_{L,k}^{\mathrm{loc}}),
eL,kloc≜[(ei1,kloc)T,…,(eiNL,kloc)T]T,(18)e_{L,k}^{\mathrm{loc}}\triangleq\left[(e_{i_1,k}^{\mathrm{loc}})^{\mathsf{T}},\ldots,(e_{i_{N_L},k}^{\mathrm{loc}})^{\mathsf{T}}\right]^{\mathsf{T}}, \tag*{(18)}

and

Jη,k=[Ci1RrkTΦ(si1,rk)⋮CiNLRrkTΦ(siNL,rk)].(19)J_{\eta,k}= \begin{bmatrix} C_{i_1}R_{rk}^{\mathsf{T}}\Phi(s_{i_1,rk})\\ \vdots\\ C_{i_{N_L}}R_{rk}^{\mathsf{T}}\Phi(s_{i_{N_L},rk}) \end{bmatrix}. \tag*{(19)}

A reference is eligible for a local magnetic update after a minimum age, minRefAgeminRefAge. It is retired when its maximum age, maxRefAgemaxRefAge, is exceeded. Eligible measurements undergo the innovation test in § 4.5. After one update attempt, accepted or rejected, the reference is retired; a new reference can be formed at a later epoch.

Global Magnetic Model and Update

The global magnetic field model represents the world-frame magnetic field. Following the reduced-rank scalar-potential Gaussian process representation used in magnetic SLAM [21], we write

BW(s)=b0+∑j=1Maj∇ϕj(s)=W(s)u,(20)B^{W}(s)=b_0+\sum_{j=1}^{M}a_j\nabla\phi_j(s)=W(s)u, \tag*{(20)}

where b0b_0 denotes the background magnetic field, u=[b0T,a1,…,aM]Tu=[b_0^{\mathsf{T}},a_1,\ldots,a_M]^{\mathsf{T}}, and W(s)=[I3,∇ϕ1(s),…,∇ϕM(s)]W(s)=[I_3,\nabla\phi_1(s),\ldots,\nabla\phi_M(s)]. The functions ϕj\phi_j are normalized Laplacian eigenfunctions on a fixed rectangular domain Ω\Omega, ordered by increasing eigenvalue λj\lambda_j. A squared-exponential potential kernel gives the prior

u∼N(0,Pu0),Pu0=blkdiag⁡(σb2I3,Λ),u\sim\mathcal{N}(0,P_u^0),\qquad P_u^0=\operatorname{blkdiag}(\sigma_b^2I_3,\Lambda),
Λjj=σp2(2πℓ2)3/2exp⁡(−ℓ2λj/2).(21)\Lambda_{jj}=\sigma_p^2(2\pi\ell^2)^{3/2}\exp(-\ell^2\lambda_j/2). \tag*{(21)}

where σb\sigma_b is the prior standard deviation of each background field component, and ℓ\ell and σp\sigma_p are the potential length scale and standard deviation. The domain is specified before processing and remains fixed during navigation. Its extent and the number of basis functions determine the spatial detail represented by the map.

For magnetometer i∈Gi \in G, the sensor position and its measurement model are

si,kW=pk+Rkdi,s^{W}_{i,k} = p_k + R_k d_i,
yi,k=hi,kGP(xk)+ei,kGP,y_{i,k} = h^{\mathrm{GP}}_{i,k}(x_k) + e^{\mathrm{GP}}_{i,k},
hi,kGP(xk)≔CiRkWW(si,kW)uk+omi.(22)h^{\mathrm{GP}}_{i,k}(x_k) \coloneqq C_i R_k^{W} W(s^{W}_{i,k})u_k + o^{m_i}. \tag*{(22)}

where ei,kGPe^{\mathrm{GP}}_{i,k} includes sensor noise and an allowance for GP approximation error. For the stacked measurements, the effective noise covariance is

ΣkGP=cov⁡(ekGP),\Sigma^{\mathrm{GP}}_k = \operatorname{cov}(e^{\mathrm{GP}}_k),
ekGP≔[(ei1,kGP)T,…,(eiNG,kGP)T]T.(23)e^{\mathrm{GP}}_k \coloneqq\left[(e^{\mathrm{GP}}_{i_1,k})^{\mathsf{T}}, \ldots, (e^{\mathrm{GP}}_{i_{N_G},k})^{\mathsf{T}}\right]^{\mathsf{T}}. \tag*{(23)}

Here NG≔∣G∣N_G \coloneqq|G| denotes the number of magnetometers in GG.

The global magnetic update is attempted at selected keyframes. The initial epoch is a keyframe. Subsequently, a new keyframe is selected after a prescribed number NaccN_{\mathrm{acc}} of local updates have been accepted since the previous keyframe.

Filter Update and Cloned-Pose Management

For innovation ν=y−h(x^)\nu= y - h(\hat{x}), measurement model Jacobian HH, and effective noise covariance Σ\Sigma,

S=HP−HT+Σ,K=P−HTS−1,δx^=Kν,P~+=(I−KH)P−(I−KH)T+KΣKT.(24)\begin{aligned} S &= H P^{-} H^{\mathsf{T}} + \Sigma, \qquad K = P^{-}H^{\mathsf{T}}S^{-1},\\ \delta\hat{x} &= K\nu,\\ \widetilde{P}^{+} &= (I-KH)P^{-}(I-KH)^{\mathsf{T}} + K\Sigma K^{\mathsf{T}}. \tag*{(24)} \end{aligned}

The update is accepted only when νTS−1ν≤χd,1−α2\nu^{\mathsf{T}}S^{-1}\nu\leq\chi^{2}_{d,1-\alpha}, where dd is the innovation dimension and α\alpha is the chosen rejection probability. Vector corrections are additive; the current and cloned attitudes are corrected by left multiplication with Exp⁡([δϕ^]×)\operatorname{Exp}([\delta\hat{\phi}]_{\times}). The covariance is then reset as P+=JresetP~+JresetTP^{+} = J_{\mathrm{reset}}\widetilde{P}^{+}J_{\mathrm{reset}}^{\mathsf{T}}, using the SO(3)SO(3) left Jacobian in each corrected attitude block.

When a pose is cloned, let EE select its position and attitude errors. The covariance augmentation is

Paug=[PPETEPEPET].(25)P_{\mathrm{aug}} = \begin{bmatrix} P & PE^{\mathsf{T}}\\ EP & EPE^{\mathsf{T}} \end{bmatrix}. \tag*{(25)}

Retiring a reference removes only the corresponding six coordinates, retaining the marginal covariance of all remaining states.

Experiments

Experimental Setup and Evaluation

We collected data using the handheld sensor rig shown in Figure 2, which integrates a nine-magnetometer array and a stereo-inertial module. The rig uses a Bosch BMI160 IMU and nine PNI RM3100 magnetometers arranged over a 181×160×87181 \times160 \times87 mm array. The stereo camera records 640×480640 \times480-pixel images at 25 Hz with a 60-mm baseline. Both the magnetometer array and stereo-inertial module share a common system clock; during acquisition, IMU readings sampled at 200 Hz were interpolated to match the 100-Hz magnetometer timestamps. In total, three datasets were recorded on two floors of a building and in an underground parking lot, as summarized in Table 1. Each recording starts and ends at the same location. The trajectories and evaluation objectives are described below.

DatasetDuration (s)Length (m)Field magnitude std. (snapshot)a^{a}(μ\muT)Field magnitude std.b^{b}(μ\muT)Visual condition
Multi-loop94901.235.38Good
Single-loop117901.5211.46Fair
Visually degraded1531202.5515.57Poor

Table 1. Characteristics of the datasets.

Handheld IMU–magnetometer-camera sensor rig with enlarged views of the three-dimensional nine-magnetometer array and the stereo-inertial module

Figure 2. Handheld IMU–magnetometer-camera sensor rig with enlarged views of the three-dimensional nine-magnetometer array and the stereo-inertial module. The magnetometer array spans 181×160×87181 \times160 \times87 mm.

Multi-loop

The three-lap square trajectory was collected on the fifth floor of a building, enabling assessment of position tracking along straight segments and heading estimation through 90∘90^\circ turns. Successive laps provide repeated visits to the same region, allowing us to assess the effect of loop closure in MISLAM.

Single-loop

The single-loop trajectory, recorded over nearly two minutes in a large hall on the sixth floor of the same building, allows us to evaluate the magnitude of the accumulated errors in exploring unknown environments.

Visually degraded

This dataset was collected in an underground parking lot with insufficient lighting. The experiment evaluates the ability of MISLAM to maintain tracking under poor illumination.

We use these datasets to compare MISLAM with MAINS, a state-of-the-art magnetic-inertial odometry [3], and OpenVINS, a visual-inertial odometry system [2]. Table 2 reports the MISLAM parameters used for each dataset.

SettingMulti-loopSingle-loopVisually degraded
Polynomial model degree nn22
Σfit\Sigma_{\mathrm{fit}} (μT2\mu\mathrm{T}^{2})0.012⋅I3NL0.01^{2}\cdot I_{3N_L}
minRefAge\mathit{minRefAge}40 ms
maxRefAge\mathit{maxRefAge}250 ms
Chi-square test rejection prob. αfit\alpha_{\mathrm{fit}}0.01
Length scale ℓ\ell1 m
NaccN_{\mathrm{acc}}20
Innovation rejection prob. α\alpha0.01
Dataset-specific
ΣL,kloc\Sigma^{\mathrm{loc}}_{L,k} (μT2\mu\mathrm{T}^{2})0.032⋅I3NL0.03^{2}\cdot I_{3N_L}0.032⋅I3NL0.03^{2}\cdot I_{3N_L}0.072⋅I3NL0.07^{2}\cdot I_{3N_L}
Number of basis functions MM50100300
σb\sigma_b110.8
σp\sigma_p505025
ΣkGP\Sigma^{\mathrm{GP}}_{k} (μT2\mu\mathrm{T}^{2})0.12⋅I3NG0.1^{2}\cdot I_{3N_G}0.82⋅I3NG0.8^{2}\cdot I_{3N_G}0.42⋅I3NG0.4^{2}\cdot I_{3N_G}

Table 2. MISLAM settings for the three datasets

These parameters were manually tuned based on the estimated trajectories. The local group contains all magnetometers except the central one, which is reserved for the global group. We evaluate all methods offline and use endpoint closure error as the evaluation metric.

Experimental Results

Table 3 summarizes the endpoint closure errors for the three datasets, while Figure 3 and Figure 4 show the estimated trajectories and height profiles. MISLAM achieves a lower total closure error than MAINS in all three experiments, with reductions of 78.0%, 58.9%, and 45.7% for the multi-loop, single-loop, and visually degraded datasets, respectively. Its performance relative to OpenVINS depends on the trajectory and visual conditions, as discussed below.

Figure 3. Trajectory overlays (top) and height estimates (bottom) for the multi-loop and single-loop experiments. Ellipses and error bars indicate 95% uncertainty regions for the final horizontal position estimates and 95% uncertainty intervals for the final height estimates, respectively. The magenta cross marks the starting point.

Figure 4. Visually degraded experiment. (a) Trajectory overlay (top) and height estimates (bottom). Scene images (b)–(g) correspond to numbered locations 1–6 in (a), respectively. OpenVINS lost tracking in the poorly lit corridor and terminated around 80 s.

ExperimentAlgorithmTotal (m)Horizontal (m)Vertical (m)
Multi-loopMAINS1.2480.6281.079
MISLAM0.2740.1610.223
OpenVINS0.9890.9890.025
Single-loopMAINS3.8313.8130.364
MISLAM1.5761.5680.163
OpenVINS0.8910.8830.119
Visually degradedMAINS6.4536.3181.315
MISLAM3.5020.3323.486
OpenVINS†^{\dagger}–––

Table 3. Endpoint closure errors across three experiments. †OpenVINS terminates early in the visually degraded environment.

In the multi-loop experiment, MISLAM achieves the smallest total and horizontal closure errors, at 0.274 m and 0.161 m, respectively. The corresponding errors are 1.248 m and 0.628 m for MAINS, and 0.989 m and 0.989 m for OpenVINS. As shown in Figure 3(a), the MISLAM trajectories from successive laps exhibit less lateral separation than those of MAINS. MISLAM also reduces the vertical closure error from 1.079 m to 0.223 m, although OpenVINS provides the smallest vertical error, at 0.025 m. These observations are consistent with the benefit of retaining a global magnetic map to constrain accumulated drift when revisiting previously explored regions.

In the single-loop experiment, MISLAM reduces the total closure error from 3.831 m for MAINS to 1.576 m. Most of this improvement occurs in the horizontal plane, where the error decreases from 3.813 m to 1.568 m. The trajectory in Figure 3(b) also follows the overall shape of the OpenVINS estimate more closely than the MAINS trajectory. OpenVINS nevertheless achieves the lowest total closure error, at 0.891 m. The larger MISLAM closure error relative to the multi-loop experiment is consistent with the more limited opportunities to revisit mapped regions during exploration, although the two datasets also differ in their trajectories and magnetic-field characteristics. The visually degraded experiment illustrates the ability to continue tracking when visual measurements become unreliable. OpenVINS loses tracking in the poorly lit corridor and terminates at approximately 80 s, whereas both magnetic-inertial methods complete the recording. MISLAM achieves a horizontal closure error of 0.332 m, compared with 6.318 m for MAINS, a reduction of 94.7%. However, the height plot in Figure 4(a) shows substantial drift in MISLAM: its vertical closure error is 3.486 m, exceeding the 1.315 m error of MAINS and dominating its total error of 3.502 m. Thus, the improvement in this experiment is primarily horizontal, and vertical drift remains a limitation.

Finally, Figure 5 visualizes the estimated magnetic-field magnitude relative to that of the calibration field along the trajectories. These plots provide a qualitative view of the field representation in MISLAM. The multi-loop dataset repeatedly covers approximately the same region, allowing MISLAM to reuse previously acquired field information to correct accumulated drift. The resulting pose corrections are visible as discontinuities in the estimated trajectory at global magnetic updates.

Estimated GP magnetic-field magnitude evaluated along the estimated trajectories and displayed in the horizontal plane

Figure 5. Estimated GP magnetic-field magnitude evaluated along the estimated trajectories and displayed in the horizontal plane. Colors indicate the magnitude of the final posterior mean relative to that of the calibration field. Each panel uses its own color scale.

Conclusions

This letter presents MISLAM, a tightly coupled magnetic–inertial SLAM system that uses an IMU and a nine-magnetometer array, without relying on external odometry or platform-specific motion assumptions. Offline evaluation on three datasets collected with a handheld sensor rig shows that MISLAM reduces total endpoint closure error by 45.7%–78.0% relative to MAINS, achieving a total closure error of 0.274 m in the multi-loop experiment. In the visually degraded experiment, MISLAM continued tracking throughout the recording after OpenVINS lost tracking and achieved a horizontal closure error of 0.332 m. These results support the potential of magnetic-inertial SLAM for indoor localization under poor illumination. However, the vertical closure error of 3.486 m in this experiment shows that accurate horizontal closure does not necessarily imply accurate three-dimensional localization. Future work will investigate the causes of vertical drift, improve height estimation, explore alternative magnetic-field models, and integrate the system into existing multisensor-fusion frameworks for field robots.

References

  1. [1]Kamak Ebadi, Lukas Bernreiter, Harel Biggie, Gavin Catt, Yun Chang, Arghya Chatterjee, Christopher E. Denniston, Simon-Pierre Deschênes, Kyle Harlow, Shehryar Khattak, Lucas Nogueira, Matteo Palieri, Pavel Petráček, Matěj Petřík, Andrej Reinke, Vít Krátký, Shibo Zhao, Ali-akbar Agha-mohammadi, Kostas Alexis, Christoffer Heckman, Kasra Khosoussi, Navinda Kottege, Benjamin Morrell, Marco Hutter, Fred Pauling, François Pomerleau, Martin Saska, Sebastian Scherer, Roland Siegwart, Jason L. Williams, and Luca Carlone. Present and future of SLAM in extreme environments: The DARPA SubT challenge. IEEE Trans. Robot., 40:936–959, 2024. Patrick Geneva, Kevin Eckenhoff, Woosik Lee, Yulin Yang, and Guoquan Huang. OpenVINS: A research platform for visual-inertial estimation. In Proc. IEEE Int. Conf. Robot. Autom. (ICRA), Paris, France, 2020.
  2. [3]Chuan Huang, Gustaf Hendeby, Hassen Fourati, Christophe Prieur, and Isaac Skog. MAINS: A magnetic-field-aided inertial navigation system for indoor positioning. IEEE Sensors J., 24(9):15156–15166, 2024.
  3. [4]Chuan Huang, Gustaf Hendeby, and Isaac Skog. Inertial magnetic SLAM systems using low-cost sensors, 2026.
  4. [5]Chuan Huang, Gustaf Hendeby, and Isaac Skog. Joint magnetometer-IMU calibration via maximum a posteriori estimation. IEEE Sensors J., 26(11):17672–17684, 2026.
  5. [6]Manon Kok and Arno Solin. Scalable magnetic field SLAM in 3D using Gaussian process maps. In Proc. 21st Int. Conf. Inf. Fusion (FUSION), pages 1353–1360, Cambridge, United Kingdom, July 2018.arxiv.org/abs/1804.01926
  6. [7]Manon Kok and Arno Solin. Online one-dimensional magnetic field SLAM with loop-closure detection. In Proc. IEEE Int. Conf. Multisensor Fusion Integr. Intell. Syst. (MFI), pages 1–7, 2024.arxiv.org/abs/2409.01091
  7. [8]Jian Kuang, Yan Wang, Longyang Ding, Baoding Zhou, Liping Xu, Li Cao, Lanqin He, Yunhui Wen, and Xiaoji Niu. CrowdMagMap 2.0: Crowdsourced magnetic mapping for multi-floor underground parking lot navigation. IEEE Trans. Intell. Transport. Syst., 26(11):18708–18721, 2025.DOI
  8. [9]Taylor N Lee and Aaron J Canciani. MagSLAM: Aerial simultaneous localization and mapping using Earth’s magnetic anomaly field. Navigation, 67(1):95–107, 2020.
  9. [10]Binghao Li, Thomas Gallagher, Andrew G Dempster, and Chris Rizos. How feasible is the use of magnetic field alone for indoor positioning? In Proc. Int. Conf. Indoor Positioning Indoor Navig. (IPIN), pages 1–9, 2012.DOI
  10. [11]Jiazhu Li, Jian Kuang, Yan Wang, and Xiaoji Niu. SC-EKF-MIO: Magnetic-inertial odometry based on stochastic clone extended Kalman filter. IEEE Trans. Instrum. Meas., 74:1–13, 2025.
  11. [12]Yuxin Liu, Wenchao Zhang, Lei Cao, Dongyan Wei, and Hong Yuan. FSS-EMD-MIO: robust map-free magnetic-inertial odometry via Fibonacci sphere-sampled equivalent dipole modeling for indoor navigation. Satellite Navig., 7(1):11, 2026.DOI
  12. [13]Natalia Pavlasek, Charles Champagne Cossette, David Roy-Guay, and James Richard Forbes. Magnetic navigation using attitude-invariant magnetic field information for loop closure detection. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), pages 5251–5257, Detroit, USA, 2023.arxiv.org/abs/2309.02394
  13. [14]Mark Pedley. Tilt sensing using a three-axis accelerometer. Application Note AN3461, Freescale Semiconductor, Inc., March 2013. Rev. 6.
  14. [15]Tongxing Peng, Chao Ming, Zhengpeng Yang, Huaiyan Wang, Jiyan Yu, and Xiaoming Wang. MAVIS: A prior-free magneto-visual-inertial odometry exploiting inhomogeneous magnetic fields for seamless indoor navigation. IEEE Internet Things J., pages 1–1, 2026.DOI
  15. [16]Hongming Shen, Zhenyu Wu, Wei Wang, Qiyang Lyu, Huiqin Zhou, and Danwei Wang. IDF-MFL: Infrastructure-free and drift-free magnetic field localization for mobile robot. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), pages 2278–2285, 2024.arxiv.org/abs/2411.06182
  16. [17]Isaac Skog, Gustaf Hendeby, and Fredrik Gustafsson. Magnetic odometry - a model-based approach using a sensor array. In Proc. Int. Conf. Inf. Fusion (FUSION), pages 794–798, Cambridge, United Kingdom, July 2018.DOI
  17. [18]Isaac Skog, Gustaf Hendeby, and Felix Trulsson. Magnetic-field based odometry – an optical flow inspired approach. In Proc. Int. Conf. Indoor Positioning Indoor Navig. (IPIN), pages 1–8, Lloret de Mar, Spain, Nov. 2021.DOI
  18. [19]Isaac Skog, Manon Kok, Gustaf Hendeby, Chuan Huang, and Thomas Edridge. On the connection between magnetic-field odometry aided inertial navigation and magnetic-field SLAM. In Proc. IEEE/ION Position Location Navig. Symp. (PLANS), pages 809–814, 2025.arxiv.org/abs/2503.04286
  19. [20]Joan Solà. Quaternion kinematics for the error-state Kalman filter, 2017.
  20. [21]Arno Solin, Manon Kok, Niklas Wahlström, Thomas B Schön, and Simo Särkkä. Modeling and interpolation of the ambient magnetic field by Gaussian processes. IEEE Trans. Robot., 34(4):1112–1127, 2018.arxiv.org/abs/1509.04634
  21. [22]Ilari Vallivaara, Yinhuan Dong, Bingnan Duan, and Tughrul Arslan. Saying goodbyes to rotating your phone: Magnetometer calibration during SLAM—extended version. IEEE J. Indoor Seamless Position. Navig., 4:41–52, 2026.DOI
  22. [23]Ilari Vallivaara, Janne Haverinen, Anssi Kemppainen, and Juha Röning. Simultaneous localization and mapping using ambient magnetic field. In Proc. IEEE Conf. Multisensor Fusion Integr. (MFI), pages 14–19, 2010.DOI
  23. [24]Frida Viset, Rudy Helmons, and Manon Kok. An extended Kalman filter for magnetic field SLAM using Gaussian process regression. Sensors, 22(8), 2022.DOI
  24. [25]David Vissiere, Alain Martin, and Nicolas Petit. Using distributed magnetometers to increase IMU-based velocity estimation into perturbed area. In Proc. 46th IEEE Conf. Decis. Control (CDC), pages 4924–4931, New Orleans, LA, USA, Dec. 2007.DOI
  25. [26]Tisheng Zhang, Linfu Wei, Jian Kuang, Hailiang Tang, and Xiaoji Niu. Mag-ODO: Motion speed estimation for indoor robots based on dual magnetometers. Measurement, 222:113688, 2023.DOI
  26. [27]Makia Zmitri, Hassen Fourati, and Christophe Prier. Improving inertial velocity estimation through magnetic field gradient-based extended Kalman filter. In Proc. Int. Conf. Indoor Positioning Indoor Navig. (IPIN), pages 1–7, Pisa, Italy, Sep. 2019.DOI
  27. [28]Makia Zmitri, Hassen Fourati, and Christophe Prier. Magnetic Field Gradient-Based EKF for Velocity Estimation in Indoor Navigation. Sensors, 20(20):5726, 2020.DOI

Paper details

vlnc:W_qjke5szx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents