Industry Research

Appearance

SNAIL Radar Benchmark ​

Related docs: 4D imaging radar RIO and SLAM, radar odometry and radar SLAM, radar-inertial odometry, radar-LiDAR-inertial fusion, and benchmarking metrics and datasets.

Last updated: 2026-05-09

Executive Summary ​

SNAIL Radar is a large-scale 4D radar benchmark for radar-based odometry, mapping, and place recognition. It matters because many radar SLAM papers historically evaluated on small, single-platform, fair-weather, or perception-oriented radar datasets. SNAIL Radar adds repeated traversals, multiple platforms, rain/night conditions, 4D radar, LiDAR, cameras, IMUs, GNSS/INS, and careful reference-trajectory generation.

The dataset was collected from September 2023 to February 2024 using handheld, e-bike, and SUV platforms. It covers campus roads, highway tunnels, overpasses, and repeated routes under clear, night, rain, dusk, and heavy-rain conditions.

Sensor Suite ​

The hardware pages describe:

  • 4D radars: Continental ARS548 and Oculii Eagle, depending on platform and date.
  • 3D LiDAR: Hesai Pandar XT32.
  • Stereo camera: ZED2i.
  • IMUs: ZED2i IMU, XSens MTi3DK, and GNSS/INS IMU sources depending on platform.
  • GNSS/INS: Bynav X36D.

Some earlier sequences do not include ARS548 and MTi3DK data, so benchmark comparisons should filter sequences by available sensor set.

Timing and Calibration ​

SNAIL Radar is especially useful for synchronization work:

  • Data packets are synchronized to GNSS time.
  • The paper describes a two-step process with convex-hull-based smoothing and correlation-based correction.
  • The website releases a time-offset and rotation calibration tool using FFT-based correlation.
  • The reference motion is generated by registering LiDAR scans to a TLS point-cloud map with a LiDAR-inertial sequential localizer supporting forward and backward processing.

This makes SNAIL Radar more useful for radar SLAM than a raw driving log with approximate timestamps.

Benchmark Tasks ​

Primary tasks:

  • 4D radar odometry.
  • Radar-inertial odometry.
  • Radar mapping.
  • Radar place recognition and loop closure.
  • Cross-condition repeatability in rain, night, and repeated routes.

Useful metrics:

  • ATE/RPE by platform and weather.
  • Drift per kilometer in tunnels and rain.
  • Radar place-recognition recall at distance thresholds.
  • Loop-closure precision/recall.
  • Doppler ego-velocity error.
  • Failure rate by radar model and point density.

Dynamic and Degraded Scenes ​

SNAIL Radar directly covers environmental degradation:

  • heavy rain and moderate rain,
  • night and dusk,
  • tunnels and overpasses,
  • vegetated campus roads,
  • repeated traversals for place recognition.

It does not fully cover airside edge cases such as aircraft multipath, jet blast spray, very open apron sparsity, or stand equipment churn, but it is one of the better public sources for radar localization stress testing.

Integration Readiness ​

The dataset provides SDK tools for loading, visualization, conversion to ROS1 bags and folders, plus calibration utilities and cascaded pose graph optimization software. This is a practical advantage for teams trying to compare radar odometry methods without building all dataset tooling from scratch.

Limitations ​

  • Sensor availability differs across sequences.
  • Radar characteristics differ strongly between ARS548 and Oculii Eagle.
  • Campus/highway data does not replace local validation for ports, mines, construction, or airports.
  • Reference trajectories depend on LiDAR/TLS processing, so radar-only claims should still inspect GT uncertainty.
  • The dataset is large and requires careful sequence selection for fair comparisons.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim