Radar is more resilient to adverse weather and lighting conditions than visual and Lidar simultaneous localization and mapping (SLAM). However, most radar SLAM pipelines still rely heavily on frame-to-frame odometry, which leads to substantial drift. While loop closure can correct long-term errors, it requires revisiting places and relies on robust place recognition. In contrast, visual odometry methods typically leverage bundle adjustment (BA) to jointly optimize poses and map within a local window. However, an equivalent BA formulation for radar has remained largely unexplored. We present the first radar BA framework enabled by Gaussian Splatting (GS), a dense and differentiable scene representation. Our method jointly optimizes radar sensor poses and scene geometry using full range-azimuth-Doppler data, bringing the benefits of multi-frame BA to radar for the first time. When integrated with an existing radar-inertial odometry frontend, our approach significantly reduces pose drift and improves robustness. Across multiple indoor scenes, our radar BA achieves substantial gains over the prior radar-inertial odometry, reducing average absolute translational and rotational errors by 90% and 80%, respectively.
Figures & tables
Fig. 1: Existing radar SLAM methods largely depend on frame-to-frame estimation, which inevitably accumulates drift over time. To address this limitation, we introduce a Gaussian Splatting–based radar bundle adjustment framework that jointly optimizes multi-frame poses and a dense scene representation. This joint optimization substantially reduces drift and improves overall trajectory consistency.
Fig. 2: Illustration of 3D radar data from a Multiple-Input and Multiple-Output (MIMO) radar after FFT-based signal processing. The resulting data cube spans three dimensions corresponding to range, azimuth, and Doppler velocity measurements.
Fig. 3: Comparison between visual and radar SLAM. Visual SLAM has progressed from frame-to-frame VO to BA-enabled joint pose–scene optimization. In contrast, existing radar SLAM remains predominantly frame-to-frame. Our method introduces the first radar BA for drift-resilient odometry. ρ(⋅) can be any loss function.
Fig. 4: Pipeline of RadarSplat-RIO. The frontend performs radar–inertial odometry to provide initial pose estimates. The backend refines these poses within a Gaussian Splat representation and maintains a local map using a sliding-window strategy. Finally, the proposed bundle adjustment jointly optimizes radar poses and the scene representation. Orange nodes indicate selected keyframes at different stages. More details is discussed in Sec. III-E .
Fig. 5: Range–Doppler rendering examples. Range–Doppler images are rendered from the range–azimuth image and the Doppler map computed from ego-velocity. The first and second rows show side and forward motions, yielding distinct Doppler patterns. F denotes the rendering operation.
Fig. 6: Overview of RadarSplat++ rendering and training pipeline. Given radar poses and ego-velocity, RadarSplat renders both range–azimuth and range–Doppler images, which are compared with measured radar images to compute losses for scene reconstruction and pose optimization.
Fig. 7: Data collection sensor rig setup.
Fig. 8: Visualization of camera and radar data from our indoor dataset. The radar range–azimuth images in Cartesian space are shown to provide a clearer sense of the radar observations.
Name
Description
Travel Distance (m)
Length (s)
Sequence 1
Meeting Room
42.46
141.07
Sequence 2
Small Loop
91.89
157.98
Sequence 3
Large Loop
144.59
232.74
Sequence 4
Small + Large Loop
206.12
263.19
Sequence 5
Medium + Large Loop
233.54
320.06
TABLE I: Experiment sequences details.
Scenes
Sequence 1
Sequence 2
Sequence 3
Sequence 4
Sequence 5
MRIO
Ours
MRIO
Ours
MRIO
Ours
MRIO
Ours
MRIO
Ours
Trans. (m)
0.94
0.19
7.57
0.42
22.48
2.04
11.04
1.05
12.9
0.97
Rot. ( ∘ )
5.03
1.91
15.51
2.06
32.38
5.25
17.62
2.66
21.55
3.28
TABLE II: Absolute Pose Error comparison between MRIO [ 30 ] and our method on various scenes.
Fig. 9: Comparison of estimated trajectory between MRIO [ 30 ] and proposed method.
Fig. 10: Scene reconstruction comparison between MRIO and the proposed RadarSplat-RIO. The proposed method not only greatly reduces odometry drift but also produces a dense and consistent scene reconstruction.
Method Components Ablations
Frontend
Backend
ATE
RIO
Loc. & Map.
BA
Trans. (m)
Rot ( ∘ )
✓
✓
✓
0.934
3.03
✓
10.99
18.42
✓
✓
2.83
8.23
✓
✓
18.27
25.91
TABLE III: Ablation study comparing the proposed full method, without the proposed backend, without the proposed BA, and without RIO frontend across all sequences. Metrics are Absolute Pose Error (APE) in translation (m) and rotation (deg). Lower is better.
Ablations
BA Frames Selection Strategy
RadarSplat++
Radius Spatial Window
Sliding Window
RD-image Loss
rBA=1
rBA=5
rBA=10(Selected)
N=2
N=5
N=10
N= ∞
w/
w/o
Trans. (m)
3.12
1.07
0.93
2.83
1.42
1.17
1.36
0.93
3.41
Rot. ( ∘ )
7.98
3.24
3.03
8.23
3.95
3.61
4.03
3.03
3.57
TABLE IV: Ablation study comparing different BA window selection strategies and the RadarSplat++ Range-Doppler image loss. Metrics are Absolute Pose Error in translation (m) and rotation (deg). Lower is better.
This paper introduces Dr-BA, a first-of-its-kind radar bundle adjustment (BA) framework that operates directly on 2D spinning radar intensity images. Unlike camera or lidar sensors, radar is largely unaffected by precipitation, making it a critical modality for autonomous systems that require all-weather robustness. Existing state estimation approaches using spinning radar typically extract sparse point clouds from range-azimuth-intensity measurements and apply point cloud alignment techniques to estimate vehicle motion, scene structure, or to localize within an existing map. In contrast, Dr-BA uses the full radar returns from multiple scans to jointly estimate dense maps and sensor poses. By formulating the problem as a separable optimization, we derive an efficient and general solution that decouples pose estimation from mapping. In addition to solving the BA problem, this formulation naturally extends to direct radar-only localization (DRL) within a previously built map. Dr-BA achieves state-of-the-art radar-based BA and cross-session localization performance, demonstrated on more than 200 km of on-road data across five distinct routes. Our implementation is publicly available at https://github.com/utiasASRL/dr_ba.
Daniil Lisus, Cedric Le Gentil, Timothy D. Barfoot
Robotics Institute, University of Toronto, Toronto, Canada
4D radar is increasingly attractive for robotic mapping because it provides range, azimuth, elevation, and Doppler measurements while remaining robust in adverse visibility conditions. Although recent radar and radar--inertial odometry methods have achieved promising online state estimation performance, offline global map refinement for 4D radar remains underexplored. This paper presents RAMBA, a radar bundle-adjustment framework for globally consistent 4D radar mapping. Given initial poses and radar frames from a radar--inertial odometry front-end, RAMBA jointly refines radar frame states using covariance-weighted geometric residuals, IMU preintegration factors, and radar ego-velocity constraints. The geometric residuals extend pairwise GICP to a multi-frame optimization by forming voxel-based correspondences across selected frames and weighting each residual with point covariances. To improve robustness against drift and revisits, RAMBA enforces temporal consistency during correspondence formation while explicitly supporting loop-closure constraints. Experiments on the ColoRadar and SNAIL Radar datasets show that RAMBA improves map consistency and usually enhances trajectory accuracy over radar--inertial odometry and pose-graph optimization baselines.
Jianzhu Huai, Yiwen Chen, Binliang Wang
State Key Lab of Info Engineering in Surveying, Mapping and Remote Sensing, 129 Luoyu Road, Hongshan District, Wuhan, Hubei, China
Radar sensing is increasingly used in mobile systems because it operates reliably under poor lighting, adverse weather, and privacy-sensitive settings where cameras and LiDAR often fail. However, most existing radar SLAM systems estimate motion through scan matching on discretized radar heatmaps, which breaks geometric continuity and fails to capture key radar sensing properties, often leading to unstable pose estimation and degraded mapping in regenerate or dynamically changing environments. We present DiffRadar, a real-time radar SLAM system that models radar observations as a differentiable, physics-aware Gaussian field rather than discrete scans. DiffRadar represents the scene as anisotropic Gaussian primitives and renders radar measurements in range-azimuth and Doppler-azimuth spaces through a differentiable radar forward model, enabling joint optimization of robot pose and scene structure directly from radar measurements. We implement DiffRadar on commodity FMCW radar hardware and evaluate it on both the public Radarize benchmark and a controlled stress-test suite that targets common radar SLAM failure modes, including corridor degeneracy, motion regime transitions, dynamic clutter, and long-horizon loop closures. DiffRadar achieves substantial reductions in trajectory error on the benchmark, with especially large gains under feature-poor corridor motion, while more than doubling map consistency and maintaining real-time performance at 70 FPS. These results show that modeling radar observations directly in the signal domain enables substantially more robust and consistent radar-only SLAM for mobile platforms.