RadLoc is a radar global descriptor and the SLAM that runs on it. A scan is reduced to one value per range band, averaged over azimuth, which makes it rotation invariant and small — forty floats for a place. Loops are registered by phase correlation, so no correspondences and no initial guess are needed.
cpp/radloc/ the C++ core: descriptor, place search, registration
python/radloc/ the same, for research use; checked against the C++
ros2_ws/src/
radloc_odometry/ radar odometry, publishing descriptors
radloc_slam/ pose-graph SLAM; writes a session
ltslam/ multi-session alignment; reads those sessions
docker/ everything runs in one container
docker compose -f docker/docker-compose.yml build
docker compose -f docker/docker-compose.yml run --rm validate # checks
docker compose -f docker/docker-compose.yml run --rm build # workspace
docker compose -f docker/docker-compose.yml run --rm slam # a sequenceslam takes SEQUENCE_DIR and SAVE_DIR. What it writes — the pose graph, the
descriptors and the scans — is a session, in the layout ltslam reads, so the
single- and multi-session halves meet without a converter.
MulRan KAIST_03, 3398 keyframes, 1723 loops. Odometry drifts to 98.3 m against
ground truth; with loop closure the error stays near 6 m for the whole 6.2 km.
Two halves of that run, each expressed in its own frame — the second retraces the first over about 78% of its length. 997 inter-session loops in 18 s. Measured as the distance from each keyframe of the second session to the nearest of the first, which needs no external ground truth:
| p50 | within 5 m | |
|---|---|---|
| before | 50.64 m | 8.9% |
| after | 0.59 m | 77.3% |
The ceiling matches the overlap: what has a counterpart is aligned, the rest has nothing to align to.
The descriptor is unweighted. Range weighting — counting distant bands for more — is applied once, when the coarse search key is built. It belongs to how two places are compared, not to the place itself.
Search is two-stage. A k-d tree over the leading range-weighted bands proposes candidates; the shortlist is re-ranked on the full unweighted descriptor, where the shared offset cancels.
Rotation is disambiguated. The FFT magnitude is centrally symmetric, so the log-polar stage recovers rotation only modulo 180°. Both candidates are de-rotated and the stronger translation peak wins; without this, five rotations in eight beyond ±90° come back flipped by π.
See cpp/radloc/README.md and python/README.md for the rest.