Doppler-Augmented Radar SLAM

• Doppler-Augmented Radar SLAM is a sensor fusion method that integrates Doppler velocity, range, and angular data to enhance ego-motion estimation and mapping accuracy.
• It combines multiple sensor inputs such as IMU, LiDAR, and cameras to address dynamic scenes, motion distortion, and multipath artifacts in various environments.
• Empirical studies demonstrate significant error reductions in translational and rotational metrics, underscoring its practical benefits for robust localization.

Doppler-augmented radar SLAM denotes simultaneous localization and mapping systems that treat radar radial-velocity measurements as estimation primitives rather than incidental by-products. In this class of methods, Doppler is combined with range and angular information from FMCW or 4D imaging radar, and often with IMU, LiDAR, or cameras, to improve ego-velocity estimation, scan matching, data association, dynamic-object suppression, bundle adjustment, and map consistency under adverse lighting, weather, or geometric degeneracy (Wang et al., 15 Apr 2025, Zhuang et al., 2023, Sie et al., 2023, Kung et al., 15 Apr 2026, Xiao et al., 20 Nov 2025).

1. Sensing models and the meaning of “Doppler-augmented”

A common hardware substrate is the 4D imaging mmWave radar, which measures range, azimuth, elevation or height, and Doppler radial velocity. In iRIOM, “4D” explicitly means 3D position plus radial velocity, with the Continental ARS548 producing a sparse 3D point cloud with an associated 1D Doppler component per target (Zhuang et al., 2023). Single-chip FMCW systems are also used in radar-only indoor SLAM, where preprocessing yields range–azimuth and Doppler–azimuth heatmaps or 3D point clouds with Doppler (Sie et al., 2023, Jiang et al., 17 Mar 2025). Spinning FMCW radar adds another regime: the scan is not instantaneous, so motion distortion and Doppler-induced range bias become part of the navigation model (Burnett et al., 2020).

The fundamental measurement relation in many monostatic radar SLAM formulations is the projection of ego-velocity onto the line of sight. In iRIOM, for a static world point observed in the radar frame at position $\boldsymbol{\rho}$, the measured Doppler is modeled as

$$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$

This same structure appears in RadarSplat-RIO, where sparse Doppler features satisfy $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$ with $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$, and in bistatic mmWave radio SLAM, where the Doppler of path $i$ is

$$d_k^i = \boldsymbol{v}_k^\mathsf{T} \boldsymbol{q}_{\text{AOA},k}^i.$$

Across these formulations, Doppler is a directional velocity constraint tied to scene geometry (Zhuang et al., 2023, Kung et al., 15 Apr 2026, Ge et al., 2022).

A distinctive complication arises in spinning radar. There, the Doppler effect is not only a motion cue but also a source of measurement bias. For a target at azimuth $\phi$ and sensor-frame planar velocity $(v_x,v_y)$, the radial component is $u = v_x \cos(\phi) + v_y \sin(\phi)$, and the Doppler-induced range correction is approximated as

$$\Delta r_{\rm corr} = \beta \, (v_x \cos\phi + v_y \sin\phi).$$

This led to the conclusion that both motion distortion and the Doppler effect are significant in different aspects of spinning-radar navigation, with the former more prominent than the latter (Burnett et al., 2020).

2. Estimation roles of Doppler

The oldest and still most pervasive role of Doppler in radar SLAM is single-scan ego-velocity estimation. iRIOM estimates a full 3D radar-frame ego-velocity from one scan by solving

$$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$0

using graduated non-convexity to reject moving objects and multipath, and then fuses the resulting velocity with IMU and scan-to-submap constraints in an IEKF (Zhuang et al., 2023). RadarSplat-RIO uses the same classical least-squares Doppler stage in its radar–inertial odometry frontend, where stacked radial-velocity equations are solved for $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$1 (Kung et al., 15 Apr 2026).

A second role is data association. In the parking-oriented robocentric EKF formulation, Doppler is bound to each individual feature state rather than to a scan-level residual. For a stationary object and radar sensor velocity $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$2 in the sensor frame, the radial relative velocity is modeled as

$$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$3

and matching uses 4D information $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$4 through a Mahalanobis distance. The paper attributes robust data association and filter convergence to this per-feature coupling of Doppler and feature direction (Diener et al., 9 Sep 2025).

A third role is direct inclusion in scan matching or registration likelihoods. Doppler Polar NDT extends polar-coordinate NDT to a 3D likelihood over range, angle, and Doppler,

$$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$5

with the predicted Doppler under pose increment $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$6 given by

$$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$7

This makes Doppler a third dimension of the scan-matching objective rather than a preprocessing cue (Heller et al., 2021).

A fourth role is dense multi-frame optimization. RadarSplat-RIO uses Doppler in a Gaussian-splatting bundle-adjustment backend through differentiable range–Doppler rendering, so that deviations in pose alter the predicted RD pattern and are penalized jointly with range–azimuth reconstruction. In its ablation, RD image loss reduced translational APE from approximately $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$8 to approximately $$v^d = \frac{\boldsymbol{\rho}^T }{\|\boldsymbol{\rho}\|} \; {}^{R}\mathbf{v}_{m}.$$9, while rotational error changed from approximately $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$0 to approximately $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$1, indicating that Doppler acted primarily as a translational constraint in that formulation (Kung et al., 15 Apr 2026).

3. Front-end odometry and scan-matching architectures

One branch of Doppler-augmented radar SLAM is radar–inertial odometry built around per-scan velocity estimation. iRIOM exemplifies this design: Doppler yields a robust single-scan ego-velocity estimate, static points are selected by the residual test $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$2, scan-to-submap matches are computed using a distribution-to-multi-distribution distance, and all terms are fused in an iterated EKF whose state includes IMU pose and velocity, biases, radar extrinsics, and gravity (Zhuang et al., 2023). The same paper reports that the ablation “Vel. only” was insufficiently robust, while the combination of Doppler ego-velocity, scan-to-submap matching, and IMU gave the best robustness and accuracy, indicating that Doppler is most effective when coupled to geometric mapping rather than used in isolation (Zhuang et al., 2023).

A second branch is radar-native, radar-only indoor SLAM. Radarize decouples translation and rotation: translation is estimated from Doppler–azimuth heatmaps by a ResNet18-based regressor, and rotation from range–azimuth heatmaps by a learned correlation model, after which Cartographer uses Doppler-based odometry as a prior for 2D submap matching and loop closure (Sie et al., 2023). The paper explicitly frames this as “radar-native SLAM,” arguing against treating radar as a noisy low-resolution lidar (Sie et al., 2023).

A third branch modifies scan-matching itself to better accommodate radar sensing physics. Polar NDT solves the registration problem in polar coordinates, and Doppler Polar NDT augments the likelihood with Doppler and SNR. The paper reports the ability to decrease the localization error by a factor of 3 to 5, depending on the scenario, with a negligible increase in computational complexity for PNDT and a moderate increase for DPNDT (Heller et al., 2021). In this family, Doppler is not merely used to remove moving objects; it directly shapes the Newton objective over the relative pose (Heller et al., 2021).

A fourth branch attacks the signal-processing frontend before any estimator is changed. “Digital Beamforming Enhanced Radar Odometry” replaces standard FFT-based processing by Capon-based 3D direction-of-arrival estimation, then estimates Doppler after beamforming, and feeds the resulting 3D points with Doppler into existing EKF-based and graph-based radar–inertial odometry systems (Jiang et al., 17 Mar 2025). The key claim is methodological: simply replacing the standard FFT-based processing with the proposed pipeline improves odometry accuracy, because Doppler constraints depend on accurate line-of-sight geometry (Jiang et al., 17 Mar 2025).

4. Back-end optimization, bundle adjustment, and calibration

Tightly coupled optimization backends are the natural extension of Doppler-augmented front ends. The abstract of Doppler-SLAM describes a Doppler-aided radar-inertial and LiDAR-inertial SLAM framework that integrates Doppler velocity measurements and spatial data into a tightly-coupled front-end and graph optimization back-end, introduces a Doppler-based scan-matching technique, and incorporates an online extrinsic calibration mechanism utilizing Doppler velocity and loop closure to dynamically maintain sensor alignment (Wang et al., 15 Apr 2025). Although the full paper content is not available in the supplied source, the abstract places graph optimization and online calibration at the center of the Doppler-aided formulation (Wang et al., 15 Apr 2025).

RadarSplat-RIO pushes the backend further into local bundle adjustment. Its map is a set of 3D Gaussians

$v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$3

and the backend jointly optimizes poses and map by minimizing RA and RD reconstruction losses over a local window. The full BA objective is

$v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$4

with a radius-based BA window of $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$5 performing best in the reported experiments (Kung et al., 15 Apr 2026).

Not all backends use Doppler as a direct odometry factor. Rad-GS uses 4D radar and a monocular camera within a 3D Gaussian Splatting SLAM system, but Doppler primarily supports dynamic-object detection and pose interpolation for unsynchronized images rather than direct graph constraints. Ego-velocity estimated from Doppler provides the boundary conditions of a cubic Hermite pose interpolation,

$v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$6

allowing unsynchronized images to refine the Gaussian map globally (Xiao et al., 20 Nov 2025). This establishes an important distinction: Doppler augmentation can operate at the level of motion factors, map rendering, or temporal alignment, and these uses are not interchangeable (Xiao et al., 20 Nov 2025).

5. Dynamics, multipath, artifacts, and common misconceptions

A recurrent misconception is that Doppler-augmented radar SLAM is simply radar SLAM with a moving-object filter. That characterization is incomplete. Dynamic/static classification is indeed central in several systems: iRIOM classifies points as static after GNC convergence using residuals

$v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$7

and only static points are passed to scan-to-submap matching (Zhuang et al., 2023). Rad-GS estimates ego-velocity by least squares from raw radar Doppler, labels dynamic detections when $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$8, propagates these labels through an octree to enhanced radar point clouds, and then projects them to image-space masks (Xiao et al., 20 Nov 2025). Yet other systems, notably DPNDT and RadarSplat-RIO, use Doppler as a continuous optimization variable rather than a binary gate (Heller et al., 2021, Kung et al., 15 Apr 2026).

Another misconception is that Doppler removes the need to reason about radar-specific artifacts. Radarize devotes a substantial portion of its pipeline to floor/ceiling suppression via elevation-aware beamforming and to multipath phantom removal by keeping only the first occupied cell along each azimuth ray (Sie et al., 2023). The spinning-radar study similarly shows that motion distortion and Doppler-induced range bias affect different parts of the pipeline, with motion distortion more prominent in odometry and Doppler more significant in metric localization (Burnett et al., 2020). These results imply that Doppler does not supersede motion-compensation, multipath handling, or uncertainty modeling; it adds another physically meaningful measurement channel that must itself be modeled correctly (Burnett et al., 2020, Sie et al., 2023).

A further point of contention concerns scene dynamics. Radarize assumes static scenes during operation and collected data with no moving objects, whereas iRIOM, Rad-GS, and several Doppler-factor formulations explicitly exploit Doppler inconsistency to reject dynamic objects (Sie et al., 2023, Zhuang et al., 2023, Xiao et al., 20 Nov 2025). This suggests that the practical meaning of “Doppler-augmented” varies across the literature: in some systems Doppler is a primary motion observable, in others it is a discriminator for static structure, and in others it is both.

Doppler has also been used beyond classical SLAM front ends, including non-line-of-sight detection and tracking. In “Seeing Around Street Corners,” temporal sequences of Doppler velocity and position measurements are fused to detect and track hidden objects in automotive scenes (Scheiner et al., 2019). A plausible implication is that future radar SLAM systems may increasingly treat Doppler not only as an ego-motion aid but also as a cue for separating static map structure, dynamic agents, and virtual reflections.

6. Representative systems, empirical performance, and open problems

The empirical literature shows that Doppler augmentation is not a marginal refinement. iRIOM reports, on an in-house campus building loop, APE translation of $v_d^i = (\mathbf{r}^i)^\top \mathbf{v}$9 for iRIOM versus $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$0 for EKFRIO, with corresponding closure errors of $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$1 horizontal and $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$2 vertical for iRIOM (Zhuang et al., 2023). Radarize reports averaged odometry ATE position of $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$3 versus $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$4 for milliEgo and $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$5 for RNIN, and end-to-end SLAM ATE position of $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$6 when using both Radarize odometry and Radarize scans, versus $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$7 for scan-matching-only RadarHD scans without odometry (Sie et al., 2023). RadarSplat-RIO reports average absolute translational and rotational error reductions of 90% and 80% relative to the prior radar–inertial odometry frontend across multiple indoor scenes (Kung et al., 15 Apr 2026). In parking, the robocentric EKF with Doppler and cross-matching achieves end-pose errors of $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$8 at the 63rd percentile and $\mathbf{r}^i = \mathbf{p}^i/\|\mathbf{p}^i\|$9 at the 95th percentile, while removing Doppler degrades those figures to $i$0 and $i$1, with a maximum error of $i$2 (Diener et al., 9 Sep 2025). DPNDT reports a factor of 3 to 5 reduction in localization error relative to conventional NDT, and the beamforming-enhanced frontend reports translation error decreases of 1.62%–22.07%, with an average of 13.68%, when existing Doppler-based odometry backends are fed improved 4D radar points (Heller et al., 2021, Jiang et al., 17 Mar 2025).

At the same time, the open problems are consistent across papers. Many formulations assume that the environment is static, or at least that a static background dominates, so heavy traffic, dense crowds, and severe multipath remain challenging (Zhuang et al., 2023, Sie et al., 2023, Diener et al., 9 Sep 2025). Radar scans remain sparse and noisy; vertical drift is poorly constrained by Doppler alone in iRIOM’s “Vel. only” ablation, and rotation gains from RD loss in RadarSplat-RIO are modest because IMU and RA geometry already dominate that component (Zhuang et al., 2023, Kung et al., 15 Apr 2026). Single-chip hardware imposes limited angular resolution, motivating beamforming and polar-coordinate estimators, while dense differentiable backends introduce computational overhead (Heller et al., 2021, Jiang et al., 17 Mar 2025, Kung et al., 15 Apr 2026). In spinning radar, scan-to-map and loop-closure style localization appear more sensitive to Doppler compensation than frame-to-frame odometry, which complicates backend design (Burnett et al., 2020).

The theoretical picture matches the empirical one. In bistatic mmWave radio SLAM, PCRB analysis shows that Doppler adds a positive semidefinite information term to both UE and landmark Fisher blocks, lowering PEB, HEB, CEB, and LEB as Doppler variance decreases (Ge et al., 2022). That result is consistent with the practical observation that Doppler most strongly improves translation, velocity, association, and map stability when geometry alone is ambiguous. The broad trajectory of the field therefore points toward systems in which Doppler is treated as a first-class measurement across the full stack: signal processing, frontend estimation, data association, backend optimization, calibration, and dynamic-scene reasoning (Wang et al., 15 Apr 2025, Kung et al., 15 Apr 2026, Jiang et al., 17 Mar 2025).