UWB Radar-Based SLAM: Methods & Applications

Updated 6 July 2026

UWB radar-based SLAM is a family of techniques that use ultra-wideband signals for simultaneous localization and mapping in challenging, feature-poor environments.
It integrates diverse sensing regimes—range-only, anchor-based, anchorless radar, and SAR imaging—to derive precise geometric constraints and environmental features.
Practical implementations demonstrate robust performance by fusing UWB data with LiDAR or vision systems, addressing challenges in GNSS-denied and obstructed scenarios.

Searching arXiv for recent and foundational papers on UWB radar-based SLAM to ground the article in cited literature. {"query":"UWB radar SLAM ultra-wideband arXiv anchorless range slam SAR All-UWB SLAM", "max_results": 10, "sort_by": "submittedDate"} {"tool":"search_arxiv","query":"UWB radar SLAM ultra-wideband arXiv anchorless range slam", "max_results": 10} Ultra-wideband radar-based SLAM denotes a family of simultaneous localization and mapping methods that use UWB radio as a primary exteroceptive source for metric state estimation and map construction. In the literature, the term spans several sensing regimes: anchor-based time-of-flight ranging, cooperative peer-to-peer range-only SLAM, anchorless impulse-radio UWB landmark sensing, and synthetic-aperture radar imaging. Several papers also note that anchor-based UWB ranging should be distinguished from “true radar-echo-based SLAM,” even though both are often grouped under the same umbrella in robotics discourse (Liu et al., 2024, Nguyen et al., 7 Oct 2025, Premachandra et al., 2023, Premachandra et al., 3 Oct 2025). Across these regimes, the common motivation is operation in GNSS-denied, feature-poor, smoke-filled, dusty, dark, or highly reflective environments in which optical modalities degrade.

1. Terminological scope and problem classes

A central distinction in the field is between range-based UWB SLAM and echo-based UWB radar SLAM. In cooperative range-only SLAM, the primary observations are inter-node distances. A representative formulation equips a robot with one UWB node and a 2D LiDAR, places several UWB beacons at unknown locations, and estimates the robot trajectory, beacon states, and a LiDAR occupancy grid without odometry, IMU, or control inputs (Song et al., 2018). In this setting, UWB provides geometric constraints among robot and beacons, while LiDAR contributes environmental structure.

A second regime is anchor-based ranging with fixed infrastructure. In this class, the robot or tag receives ranges to anchors at known or calibratable positions. “Range-SLAM” constructs a 2D occupancy grid from distances and RSSI with respect to known anchors and a motion prior, whereas coordinate-consistent fusion methods first calibrate anchor coordinates and then use them to anchor SLAM trajectories from later runs into a persistent global frame (Liu et al., 2024, Nguyen et al., 7 Oct 2025). These systems are infrastructure-dependent, but they provide an absolute metric reference.

A third regime is infrastructure-free onboard UWB radar SLAM. “UWB Radar SLAM: an Anchorless Approach in Vision Denied Indoor Environments” uses only onboard IR-UWB radar modules to detect natural point landmarks such as metal rods and table legs, derives range and bearing through local triangulation, and fuses them in EKF-SLAM (Premachandra et al., 2023). “All-UWB SLAM Using UWB Radar and UWB AOA” extends this idea by combining UWB radar point features with deployed UWB AoA tag landmarks in feature-deficient environments (Premachandra et al., 21 Jul 2025).

A fourth regime is UWB radar imaging for mapping, in which the robot motion synthesizes aperture. “Novel UWB Synthetic Aperture Radar Imaging for Mobile Robot Mapping” forms SAR images from side-mounted UWB radars and uses classical feature detectors on SAR keyframes for loop closure, while “Integration of UWB Radar on Mobile Robots for Continuous Obstacle and Environment Mapping” produces an obstacle point cloud from CIR peaks, filtering, and clustering as a front-end for later SLAM back-ends (Premachandra et al., 3 Oct 2025, Giurea et al., 30 Nov 2025). This suggests that “UWB radar-based SLAM” is best understood as a spectrum of estimation problems rather than a single canonical architecture.

2. Measurement primitives and sensing models

The most common primitive is the UWB range measurement. In cooperative range-only SLAM, the peer-to-peer distance between nodes $i$ and $j$ at time $k$ is modeled as

$z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$

with $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ under LOS conditions (Song et al., 2018). In anchor-based visual-UWB systems, the analogous model is

$\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$

and the paper explicitly frames UWB as the metric cue that resolves monocular scale ambiguity (Shi et al., 2019). Reported examples include $\sigma_n \approx 0.1$ m for Decawave DWM1000 cooperative ranging, $\sigma_\rho = 0.01$ m in simulated visual-UWB experiments, and two-way ranging implementations that avoid global clock synchronization (Song et al., 2018, Shi et al., 2019).

Anchorless UWB radar SLAM uses a different front-end even when the backend again consumes range-like constraints. In the anchorless IR-UWB system, time-of-flight obeys

$r = \frac{c\Delta t}{2},$

and peaks in the sampled range profile are filtered by statistical thresholding, Savitzky–Golay smoothing, peak detection, and range gating. Because the radar module does not directly provide AoA, bearing is derived from two collocated monostatic sensors using geometric triangulation, yielding measurements $z_t^i = [r_t^i,\phi_t^i]$ for EKF-SLAM (Premachandra et al., 2023). In the mobile obstacle-mapping front-end based on IEEE 802.15.4/802.15.4z hardware, the CIR amplitude and STS phase are used to estimate distance and PDoA-based AoA, with the accepted angular sector restricted to $j$ 0 (Giurea et al., 30 Nov 2025).

A further extension is UWB AoA anchor-tag sensing. In All-UWB SLAM, on-board anchors arranged in a ring estimate range $j$ 1 and bearing $j$ 2 to deployed tags, and the transformed tag observation in robot coordinates is obtained through a chained homogeneous transform. The same paper also gives a dead-zone radius

$j$ 3

making explicit that sensing geometry, field of view, and anchor-ring radius directly constrain observability (Premachandra et al., 21 Jul 2025).

Synthetic-aperture formulations treat each radar return as a sampled waveform rather than an immediate landmark observation. In mobile UWB SAR imaging, matched filtering is followed by back-projection, with per-pixel slant range

$j$ 4

and ideal range resolution

$j$ 5

For the reported LT102 setup with $j$ 6 GHz, the paper gives $j$ 7 cm and a range-bin spacing of approximately $j$ 8 mm from the sampling frequency $j$ 9 GHz (Premachandra et al., 3 Oct 2025).

3. Estimation architectures

The field contains several recurrent estimation patterns, but they differ mainly in how UWB information is injected into the state estimator. A tightly coupled filtering-and-refinement architecture is exemplified by cooperative UWB/LiDAR fusion. There, an EKF first estimates positions and velocities of the robot and all beacons from pairwise ranges, with state

$k$ 0

and a constant-velocity motion model. A second stage then performs joint scan matching and range re-fitting, solving for an offset $k$ 1 that corrects robot pose, heading, and beacon positions through a Gauss–Newton step, after which the corrected state feeds back into the EKF and occupancy update (Song et al., 2018). This architecture is explicitly described as a two-stage, tightly coupled loop.

Factor-graph smoothing is prevalent when UWB is fused with vision or multi-sensor odometry. The visual-UWB system formulates a joint nonlinear optimization on a Lie manifold over vehicle poses $k$ 2, visual landmarks, and anchor positions, with cost

$k$ 3

and implements tracking, local mapping, and loop closing threads in g2o following an ORB-SLAM-style design (Shi et al., 2019). VR-SLAM uses a multi-stage design instead: a UWB-aided 7-DoF global alignment module initializes visual odometry in the anchor frame via QCQP or NLS, and later UWB-aided bundle adjustment and pose graph optimization control short-term scale error and long-term drift (Nguyen et al., 2023).

Distributed multi-robot settings require a different decomposition. The distributed ranging SLAM formulation uses short bursts of inter-robot UWB ranges and odometry to estimate pairwise relative poses, rejects suspicious loop closures via Pairwise Consistency Maximization, and then fuses inlier constraints by Distributed Pose Graph Optimization. Its batch relative-pose estimation over a time window minimizes

$k$ 4

with a coarse polar search followed by nonlinear least squares (Liu et al., 2022). The emphasis here is not environmental mapping but distributed trajectory recovery in perceptually degraded spaces.

A related architectural pattern is delayed UWB integration after calibration. VIRAL SLAM first runs a camera–IMU–lidar odometry thread and a global BA thread, estimates the transform between the UWB anchor frame and the SLAM frame together with a static UWB range bias, and only then injects UWB range factors into the sliding-window odometry (Nguyen et al., 2021). The continuous-time calibration and fusion framework for coordinate-consistent localization uses a similar two-stage logic: one full run calibrates 3D anchor positions and biases by batch optimization with anchor-to-anchor distance factors and height priors, and subsequent runs use a sliding-window optimizer that fuses SLAM odometry and UWB ranges directly in the calibrated anchor frame (Nguyen et al., 7 Oct 2025). This suggests that UWB often plays a dual role: it is both a measurement source and a global reference mechanism.

4. Map representations and environmental inference

The map representation in UWB radar-based SLAM is tightly coupled to the information content of the sensing front-end.

Regime	Measurements	Map output
Cooperative UWB/LiDAR fusion	Peer-to-peer UWB ranges and 2D LiDAR scans	LiDAR-based occupancy grid map
Anchor-based Range-SLAM	Distances and RSSI to known anchors	2D occupancy grid $k$ 5
Anchorless IR-UWB EKF-SLAM	Peak-based range and triangulated bearing	Natural point landmarks
All-UWB SLAM	Radar point features and UWB AoA tag observations	Point features and tag landmarks
Mobile UWB radar front-end	CIR peaks, PDoA, clustering	Obstacle point cloud
UWB SAR imaging	Matched-filtered returns and back-projection	SAR image keyframes for mapping and loop closure

Occupancy-grid methods appear when a second modality or a ray-based occupancy hypothesis is available. In cooperative UWB/LiDAR fusion, LiDAR scan endpoints are transformed into the UWB-defined coordinate frame and matched against an occupancy probability field $k$ 6 using HectorSLAM-style Gauss–Newton alignment; the corrected scan is then inserted into a multi-resolution occupancy grid (Song et al., 2018). In Range-SLAM, the map is generated by raycasting between the current tag pose and anchors. LOS rays accumulate free evidence, NLOS rays accumulate occupied evidence, and the binary map is obtained by a sign function after temporal accumulation and filtering (Liu et al., 2024).

Landmark maps appear when the front-end can isolate sparse, repeatable reflectors. The anchorless IR-UWB EKF-SLAM system clusters provisional observations with DBSCAN over a sliding window and injects cluster centroids as point-landmark measurements. The resulting map consists of natural point landmarks rather than dense surfaces (Premachandra et al., 2023). All-UWB SLAM retains this landmark-based character but supplements natural radar point features with uniquely identified AoA tag landmarks deployed in feature-deficient segments, explicitly treating the tags as artificial landmarks for loop closure and drift correction (Premachandra et al., 21 Jul 2025).

Point-cloud and image representations arise in systems that retain more of the raw radar signal structure. The mobile UWB obstacle-mapping front-end transforms filtered CIR/PDoA detections into Cartesian points, clusters them with DBSCAN, and produces a point cloud that is intended for later EKF-SLAM or graph-SLAM back-ends (Giurea et al., 30 Nov 2025). Mobile UWB SAR imaging instead forms local SAR images by matched filtering and back-projection, then uses feature matching across SAR keyframes to generate loop-closure constraints. This is not an occupancy grid in the usual LiDAR sense; it is a synthetic-aperture intensity map whose geometry is tied to the robot path and imaging aperture (Premachandra et al., 3 Oct 2025).

The diversity of map outputs is therefore not incidental. It reflects a deeper partition of the field into methods that infer occupancy from radio geometry, methods that localize sparse reflectors as landmarks, and methods that reconstruct radar images from the accumulated aperture.

5. Observability, robustness, and recurrent failure modes

Observability is a first-order issue in nearly every UWB SLAM formulation. In cooperative range-only SLAM, the geometry is determined only up to translation and rotation, so one beacon is fixed at the origin and another on the positive $k$ 7-axis to define the local UWB frame (Song et al., 2018). In visual-UWB systems, monocular scale ambiguity is unobservable from images alone and becomes metric only after range factors are added; VR-SLAM formalizes this through a 7-DoF global alignment problem and an explicit Fisher Information Matrix analysis showing that at least seven scalar range measurements are required for full 7-DoF observability (Shi et al., 2019, Nguyen et al., 2023).

The literature repeatedly connects observability to geometry. Cooperative UWB/LiDAR fusion notes that non-trivial network geometry, multiple pairwise measurements, and LOS connectivity increase the rank of the UWB information matrix, whereas collinear or sparse configurations degrade consistency (Song et al., 2018). Visual-UWB navigation recommends anchor deployment with good geometry and low DOP, for example by dropping anchors when far from existing anchors, while VIRAL SLAM delays UWB integration until the global graph satisfies a spatial-diversity condition based on the singular values of a covariance-like matrix over keyframe positions (Shi et al., 2019, Nguyen et al., 2021). The continuous-time calibration framework likewise uses anchor-to-anchor distance factors and height priors to mitigate poor 3D geometry in cross-session localization (Nguyen et al., 7 Oct 2025).

A common misconception is that UWB robustness to smoke or darkness implies immunity to NLOS and multipath. The literature does not support that interpretation. Cooperative range-only SLAM removes suspected NLOS ranges using a channel-power heuristic: received power minus first-path power greater than $k$ 8 dB indicates likely NLOS (Song et al., 2018). Range-SLAM uses an SVM over $k$ 9, with normalization, outlier rejection, and smoothing, to derive a soft weight $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 0 for NLOS-aware WLS and map updates (Liu et al., 2024). Distributed multi-robot ranging SLAM relies on PCM to prune outlier loop closures caused by noisy or biased ranges (Liu et al., 2022). Anchorless radar landmark systems use Mahalanobis gating and DBSCAN to suppress multipath-induced spurious peaks, and All-UWB SLAM adds motion-gated collection, clustering, and RSSI heuristics to reject ghost AoA detections (Premachandra et al., 2023, Premachandra et al., 21 Jul 2025).

Several limitations recur across paradigms. UWB ranging is typically less accurate than LiDAR, so weighting between modalities is critical; the cooperative fusion paper explicitly notes an order-of-magnitude difference, roughly $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 1 cm for UWB versus $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 2 cm for LiDAR, and reports that heading derived from robot velocity becomes unreliable when the robot stops (Song et al., 2018). All-UWB SLAM reports empirical AoA errors that can exceed datasheet $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 3, as well as a dead zone near the robot induced by anchor-ring geometry (Premachandra et al., 21 Jul 2025). The mobile UWB obstacle front-end reports near-field failure below roughly $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 4 cm because the object peak merges with the direct path, and AoA reliability is limited outside $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 5 (Giurea et al., 30 Nov 2025). SAR mapping, finally, depends on short-segment odometry accuracy because aperture synthesis assumes sufficiently accurate radar poses during back-projection (Premachandra et al., 3 Oct 2025). Taken together, these results suggest that robustness in UWB SLAM is achieved by explicit handling of geometry, calibration, and radio pathologies rather than by the sensing modality alone.

6. Empirical performance, operating regimes, and research directions

Empirical evaluations show that UWB-based SLAM can be competitive in operating regimes that are difficult for optical systems. In a $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 6 m $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 7 workshop with five beacons and robot speed around $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 8 m/s, cooperative UWB/LiDAR fusion with $z_{r,k}^{ij} = \|p_i - p_j\| + n_{r,k},$ 9 and correction reduced average beacon error to $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 0 m with standard deviation $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 1 m, achieved drift-free mapping on a trajectory where HectorSLAM alone drifted, and reconstructed a $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 2 m corridor whose length HectorSLAM severely underestimated to about $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 3 m in a feature-poor case (Song et al., 2018). In the visual-UWB exploration system evaluated on EuRoC MAV MH_03_medium, the reported absolute position error mean is approximately $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 4 m, scale error approximately $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 5, average anchor error approximately $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 6 m across five anchors, and relocalization translation error approximately $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 7 m (Shi et al., 2019).

Infrastructure-free onboard radar results are also strong in small indoor spaces. The anchorless IR-UWB EKF-SLAM system reports RMS absolute trajectory error of $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 8 mm in a cluttered lab and $n_{r,k} \sim \mathcal{N}(0,\sigma_n^2)$ 9 mm in a classroom after ICP alignment to ground truth, while preserving operation in dense smoke where LiDAR and camera observations fail (Premachandra et al., 2023). All-UWB SLAM, which combines radar point features with deployed AoA tags, reports RMS ATE of $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 0 cm on a U-shaped path in smoke and $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 1 cm in a larger environment with four deployed tags, while ablations show that AOA-only and radar-only variants are less reliable (Premachandra et al., 21 Jul 2025).

Anchor-based smoke-resistant mapping also reports centimeter-scale localization. In a $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 2 m real-world setup with four anchors, Range-SLAM gives ATE RMSE of approximately $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 3 cm in the multi-agent case, $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 4 cm in the dual-agent case, and $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 5, $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 6, and $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 7 cm on three single-agent paths. The paper further reports successful operation at visibility of approximately $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 8 m in a smoke-filled scenario (Liu et al., 2024). For multi-robot featureless environments, distributed UWB+odometry SLAM reports $\rho_{k,j} = \|p_b^w(k)-p_{a_j}^w\| + n_\rho,$ 9 m translation RMSE and $\sigma_n \approx 0.1$ 0 rotation RMSE, outperforming both a centralized visual-inertial-range baseline and pure odometry in the reported indoor experiments (Liu et al., 2022).

Front-end mapping studies indicate that UWB radar can already provide usable environmental evidence before a full SLAM backend is added. The mobile obstacle-mapping pipeline based on CIR peak detection, PDoA filtering, and clustering reports obstacle-detection precision of at least $\sigma_n \approx 0.1$ 1 and recall of $\sigma_n \approx 0.1$ 2 on channel 9, with median distance error $\sigma_n \approx 0.1$ 3 cm and $\sigma_n \approx 0.1$ 4 of errors below $\sigma_n \approx 0.1$ 5 cm (Giurea et al., 30 Nov 2025). The mobile UWB SAR imaging paper reports that its SAR map differs from a LiDAR-based occupancy map in approximately $\sigma_n \approx 0.1$ 6 of cells and that AKAZE and ORB provide complementary loop-closure performance on SAR regions (Premachandra et al., 3 Oct 2025). For globally consistent cross-session localization, the continuous-time calibration and fusion framework reports that eleven of twelve fused sequences achieve ATE below $\sigma_n \approx 0.1$ 7 m after anchors are calibrated from a single run (Nguyen et al., 7 Oct 2025).

Several research directions recur across the literature. Cooperative ranging papers call for stronger NLOS mitigation, explicit robust costs, and better extrinsic and time calibration (Song et al., 2018). Range-SLAM points toward autonomous navigation, direct use of true UWB impulse radar echoes, and 3D mapping (Liu et al., 2024). All-UWB SLAM emphasizes optimal deployment of artificial landmarks, improved AoA calibration, and autonomous exploration strategies in feature-deficient spaces (Premachandra et al., 21 Jul 2025). Front-end mapping work highlights adaptive thresholds, learning-based peak classification, Doppler use, and integration with standard EKF- or graph-SLAM back-ends (Giurea et al., 30 Nov 2025). A plausible implication is that the field is converging toward hybrid systems in which UWB ranging, UWB radar echoes, AoA, and conventional inertial or wheel odometry are fused within robust optimization frameworks, with the exact map representation determined by how much environmental structure the radio front-end can reliably expose.