Industry Research

Appearance

4D Imaging Radar RIO and SLAM ​

Related docs: 4D imaging radar, radar FMCW/MIMO/Doppler, radar ambiguity and Doppler limits, radar-inertial odometry, radar odometry and radar SLAM, radar-LiDAR-inertial fusion, RadarSplat RIO, and robust multi-sensor localization.

Executive Summary ​

4D imaging radar RIO/SLAM uses radar detections with range, azimuth, elevation, Doppler, and often intensity/RCS, fused with IMU measurements. Compared with older 2D scanning radar or sparse automotive radar, 4D imaging radar adds vertical structure and denser point clouds, making radar-inertial mapping more plausible.

Representative systems include iRIOM, Go-RIO, RIO-Vehicle, and x-RIO/multi-radar yaw-aiding work. iRIOM introduced a submap-based 4D radar-inertial odometry and mapping pipeline with robust Doppler ego-velocity, scan-to-submap matching, an iterated EKF, and loop closure. Go-RIO adds ground-optimized radar filtering and continuous velocity preintegration with Gaussian processes to better handle asynchronous radar/IMU streams and ground-vehicle assumptions.

For AV and airside autonomy, this is one of the most important adverse-weather localization directions. Radar can continue operating through fog, dust, smoke, rain, spray, darkness, and glare. The caveat is equally important: radar is noisy, sparse, multipath-prone, and dynamic-object-sensitive. RIO should be a robust aiding layer, not an unchecked sole authority.

Problem Fit ​

4D imaging radar RIO fits:

  • Fog, rain, dust, smoke, snow, spray, darkness, and glare.
  • Open outdoor environments where cameras and LiDAR can degrade.
  • Ground vehicles that can exploit nonholonomic and ground-plane assumptions.
  • Long-term localization where radar reflectors and static structures remain visible.
  • Safety fallback when LiDAR/camera health is poor.

It is weaker for:

  • Dense indoor metal environments with severe multipath.
  • Scenes dominated by moving objects.
  • Very open spaces with few radar reflectors.
  • Applications requiring dense geometric maps comparable to LiDAR.

Sensor Model ​

4D radar detections typically include:

text
z_i = [range_i, azimuth_i, elevation_i, doppler_i, rcs_i, t_i]

Converted to radar-frame points:

text
p_i^R = range_i * [cos(el) cos(az), cos(el) sin(az), sin(el)]

Doppler radial velocity constrains relative motion:

text
v_d,i = u_i^T ( v_R + omega_R x p_i^R ) + noise

The inertial state is:

text
x = [R, p, v, b_g, b_a, g, T_RI, delta_t_RI]

where T_RI is radar-IMU extrinsic calibration and delta_t_RI is the time offset when estimated or modeled.

Maps may be:

  • Local radar point submaps.
  • Reflectivity/intensity submaps.
  • Gaussian or distributional radar maps.
  • Pose graphs with loop closure.

Pipeline ​

  1. Radar preprocessing

    • Filter by range, RCS, elevation, Doppler validity, and sensor-specific quality.
    • Reject ego-vehicle reflections, multipath, and dynamic-object candidates.
    • Model ground and remove unreliable ground points when useful.

  2. Doppler ego-velocity

    • Use static returns to estimate body velocity.
    • Apply RANSAC, GNC, M-estimation, or robust filtering to reject moving objects.

  3. IMU propagation

    • Propagate pose, velocity, and biases at IMU rate.
    • Deskew radar points or integrate radar velocity asynchronously.

  4. Radar scan-to-map matching

    • Register current 4D radar points to local submaps.
    • Use point-to-distribution, distribution-to-distribution, NDT/GICP, or radar-specific distances.

  5. Fusion

    • iRIOM-style systems fuse ego-velocity and scan-to-submap matches in an iterated EKF.
    • Go-RIO-style systems use continuous velocity integration and Gaussian-process interpolation.
    • Factor-graph variants add wheel, vehicle, GNSS, loop closure, or multi-radar yaw constraints.

  6. Loop closure and mapping

    • Detect revisits using radar place recognition or scan descriptors.
    • Optimize pose graph/submaps to reduce drift.

Mathematical Mechanics ​

Robust Doppler velocity estimation:

text
v_R* = arg min_v sum_i rho( z_d,i - u_i^T v_R )

with angular velocity/lever-arm extension:

text
z_d,i = u_i^T ( v_R + omega_R x p_i^R )

iRIOM-style fusion:

text
x_k^- = propagate_imu(x_k-1, u_imu)
x_k^+ = IEKF_update(x_k^-, r_doppler, r_scan_to_submap)

Scan-to-submap matching can be represented as:

text
r_scan = D( T_k p_i, M_submap )

where D is a radar-robust distributional distance rather than simple nearest-neighbor Euclidean distance.

Go-RIO emphasizes continuous velocity preintegration:

text
Delta p_ij = integral_i^j R(t) v_radar(t) dt

with Gaussian-process interpolation to align asynchronous IMU and radar velocity observations. This addresses a common RIO weakness: discretized propagation can misuse radar velocity when timestamps and rates differ.

A generic factor graph:

text
X* = arg min_X
    sum_imu      || r_imu ||^2
  + sum_dopp     rho(|| r_doppler ||^2)
  + sum_rscan    rho(|| r_radar_scan ||^2)
  + sum_vehicle  || r_nonholonomic ||^2
  + sum_loop     rho(|| r_loop ||^2)

Assumptions ​

  • A sufficient fraction of radar returns are static.
  • Radar-IMU extrinsics and timing are known or estimated.
  • Doppler ambiguity and velocity wrapping are handled.
  • Radar elevation estimates are accurate enough for 3D matching.
  • Multipath and ghost detections are robustly rejected or downweighted.
  • Ground vehicle constraints are valid when used.
  • Radar submaps are not polluted by moving vehicles or temporary objects.

Strengths ​

  • Strong adverse-weather and low-light robustness.
  • Doppler gives direct velocity information in a single scan.
  • 4D radar adds vertical structure absent in 2D scanning radar.
  • IMU closes short-term gaps and stabilizes orientation.
  • Submaps and loop closure can reduce drift.
  • Multi-radar configurations can improve yaw and coverage.
  • Good complement to LiDAR/camera localization in safety stacks.

Limitations ​

  • Radar point clouds are noisy and sparse relative to LiDAR.
  • Multipath is severe near metal, wet ground, glass, fences, and aircraft.
  • Dynamic objects violate static Doppler assumptions.
  • Yaw and position can drift without good spatial radar structure.
  • Open areas may have too few stable returns.
  • Radar maps are less geometrically interpretable than LiDAR maps.
  • Sensor-specific signal processing affects transfer across radar models.

Datasets and Benchmarks ​

Relevant datasets:

  • iRIOM author and third-party datasets: used for 4D radar-inertial mapping evaluation.
  • Go-RIO datasets: 4D radar-inertial experiments released with the method.
  • Coloradar: 4D radar, LiDAR, camera, IMU, and ground truth resources.
  • Boreas: radar/LiDAR/camera/IMU/GNSS across seasons.
  • Oxford Radar RobotCar: long-term radar driving, primarily scanning radar.
  • MulRan: radar/LiDAR urban data for odometry and place recognition.
  • K-Radar: 4D radar dataset useful for perception and radar robustness context.
  • Custom airside radar data: required for aircraft multipath, terminal reflections, rain, spray, and open apron sparsity.

Metrics:

  • ATE/RPE by weather and scene type.
  • Velocity RMSE and Doppler inlier rate.
  • Yaw drift.
  • Scan-to-submap residual distribution.
  • Loop closure precision/recall.
  • Map consistency and ghost-object contamination.
  • Availability under LiDAR/camera degradation.
  • Runtime P95/P99 on embedded compute.

AV Relevance ​

4D imaging radar RIO is highly relevant for AVs because it supplies a physically different localization signal. It can continue when cameras and LiDAR are degraded and can directly measure radial velocity.

Production use should include:

  • Wheel odometry and nonholonomic constraints.
  • LiDAR and camera factors when healthy.
  • GNSS/RTK with multipath gating.
  • HD map or radar-map localization.
  • Per-modality health and covariance inflation.
  • Dynamic-object filtering shared with perception.

The strongest production pattern is not radar replacing LiDAR. It is radar preserving observability when LiDAR/camera constraints are weak.

Indoor/Outdoor Relevance ​

Indoor: Useful in smoke, dust, darkness, tunnels, mines, warehouses, and hangars, but multipath is a major risk.

Outdoor: Strong fit for roads, ports, airports, mines, construction, and agriculture.

Mixed indoor/outdoor: Useful for transitions through hangars, underpasses, and terminal edges where lighting/GNSS/LiDAR conditions change.

Airside Deployment Notes ​

Airside is one of the strongest fits:

  • Radar works in fog, rain, night, and de-icing spray.
  • Terminal walls, signs, poles, fences, service vehicles, and gate infrastructure provide radar reflectors.
  • Aircraft provide strong returns but also multipath and dynamic/non-map geometry.
  • Open apron zones may need artificial reflectors or surveyed radar landmarks.

Recommended architecture:

  • Use RIO as an adverse-weather odometry factor.
  • Fuse wheel and nonholonomic constraints for low-speed GSE.
  • Gate aircraft returns as dynamic or map-transient unless explicitly modeled.
  • Monitor Doppler inlier geometry and multipath indicators.
  • Use surveyed radar/LiDAR map localization for global pose where available.

Validation Checklist ​

  • Calibrate radar-IMU extrinsics and time offset.
  • Validate Doppler sign convention, mounting lever arm, and velocity scale.
  • Test moving-object contamination with aircraft, tugs, buses, carts, and pedestrians.
  • Test multipath near aircraft fuselage, terminal glass/metal, fences, and wet tarmac.
  • Compare Doppler-only, scan-matching-only, and fused modes.
  • Log inlier geometry, residuals, covariance, and degradation state.
  • Test radar sensor blockage and water film effects.
  • Validate loop closures under route repeats and seasonal/weather changes.
  • Measure runtime and latency with full radar point rate.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim