Industry Research

Appearance

Radar Teach-Repeat Localization ​

Summary ​

Radar teach-repeat localization lets a robot record a route during a teaching pass and later repeat that route using radar-based localization against the taught experience. It is route-following localization, not general radar odometry. Radar odometry estimates frame-to-frame motion from live radar scans; radar teach-repeat localizes the live robot to a stored route graph or taught keyframes and uses that relative pose to follow the route.

Recent systems include Radar Teach and Repeat, CFEAR-Teach-and-Repeat, and cross-modal LiDAR Teach, Radar Repeat. VT&R3 provides the broader teach-and-repeat software architecture with radar and radar-LiDAR pipeline support. CFEAR-TR is the freshest radar-only result in this line, with an ICRA 2026 arXiv record reporting held-out Boreas localization as low as 0.117 m and 0.096 degrees at 29 Hz.

What It Is ​

Teach-and-repeat navigation separates route learning from route execution:

  • Teach: drive a route while building a graph of keyframes, path geometry, and local sensor data.
  • Repeat: localize against the taught graph and track the route without requiring global GNSS or a full global HD map.

Radar teach-repeat replaces or augments the usual camera/LiDAR localization sensor with radar. It is especially relevant when the repeat phase must work in dust, fog, smoke, rain, snow, darkness, or vegetation/scene changes that degrade optical sensors.

Core Idea ​

The system stores radar experiences along a route, then localizes each live radar scan against nearby taught keyframes and/or recent live keyframes. CFEAR-TR uses sparse oriented surface points from Doppler-compensated spinning radar measurements and aligns live scans jointly to stored scans and a sliding window of live keyframes. Radar Teach and Repeat integrates radar localization into a full closed-loop route-following system. Cross-modal LiDAR Teach, Radar Repeat teaches with precise LiDAR structure and repeats with 4D millimeter-wave radar for degraded conditions.

The important distinction from radar odometry:

  • Radar odometry can drift indefinitely because it integrates live scan-to-scan motion.
  • Radar teach-repeat constrains motion to a taught route and estimates relative pose to stored route experiences.
  • It may use radar odometry internally, but the navigation objective is repeatable path following.

Inputs and Outputs ​

ItemRole
Taught route graphKeyframes, path geometry, and traversal topology.
Stored radar scans or radar featuresLocalization reference during repeat.
Live radar scansCurrent observations for route-relative localization.
IMU/gyro/wheel odometryShort-term motion prior and controller support.
Optional LiDAR teach mapUsed in cross-modal LiDAR-teach/radar-repeat systems.
Relative pose to routeLateral, longitudinal, yaw, and sometimes full SE(3) correction.
Path-tracking commandSteering/speed command or pose target for the controller.
Localization healthMatch score, inlier count, route ambiguity, and covariance.

Pipeline ​

  1. Teach pass

    • Drive the intended route manually or under supervised autonomy.
    • Record radar scans, odometry, IMU, route topology, and controller-relevant path geometry.

  2. Experience graph construction

    • Select keyframes and store local radar representations.
    • Add odometric edges, loop closures, and route branches if the framework supports a network.

  3. Repeat initialization

    • Start near a known route segment or use retrieval to identify the current place.
    • Initialize the live robot relative to the taught graph.

  4. Radar localization

    • Align live radar scans to nearby taught keyframes.
    • Optionally align to recent live keyframes for short-term consistency.
    • Reject ambiguous or low-inlier matches.

  5. Route tracking

    • Convert route-relative pose into path-following commands.
    • Use gyro/wheel/IMU for smooth control between radar updates.

  6. Experience management

    • Add new experiences when route appearance changes.
    • Retain enough seasonal/weather diversity to avoid overfitting one teach run.

Strengths ​

  • Works without global GNSS on previously taught routes.
  • Radar is robust to darkness, dust, smoke, fog, rain, and some vegetation/appearance changes.
  • Route-relative localization can be easier than full global map localization.
  • Taught graph limits search space and reduces false global matches.
  • CFEAR-style oriented surface points make radar scans more registration-friendly.
  • The CFEAR-TR paper reports efficient single-spinning-radar repeat localization on held-out Boreas sequences, narrowing the heading-estimation gap to LiDAR-style route localization.
  • Cross-modal teaching can exploit high-quality LiDAR in good conditions and radar in degraded repeat conditions.

Failure Modes ​

  • The robot must remain on or near taught routes; it is not an open-world planner by itself.
  • Open areas with weak radar reflectors can produce poor route-relative constraints.
  • Specular radar landmarks can change with approach angle, wet surfaces, metal objects, and multipath.
  • Repeated gates, service roads, rows of poles, or similar off-road corridors can cause route aliasing.
  • Dynamic objects can dominate radar returns during repeat.
  • Structural changes along the route can invalidate stored experiences.
  • A route taught in one sensor configuration may not transfer cleanly after radar firmware, mounting, or calibration changes.
  • Cross-modal LiDAR-teach/radar-repeat needs careful representation alignment between LiDAR structure and radar observability.

Airside/AV Fit ​

Radar teach-repeat is well matched to fixed operational routes: depot-to-stand routes, baggage loops, perimeter roads, snow-removal paths, and service corridors. Airports often need repeatable path following in adverse weather and at night, and a taught-route system can reduce dependency on global GNSS in terminal-adjacent multipath.

Airside recommendations:

  • Teach routes in multiple stand occupancy states and weather conditions.
  • Avoid relying on aircraft surfaces as route landmarks.
  • Add fixed radar-observable landmarks where open-apron structure is insufficient and operations allow it.
  • Use route health metrics to slow or stop when radar localization is weak.
  • Keep route graphs versioned with airport construction and operational changes.
  • Fuse radar teach-repeat with wheel odometry, IMU, LiDAR/GNSS when healthy, and geofenced route constraints.

For road AVs, radar teach-repeat is most useful for constrained routes, depots, mines, ports, campuses, agriculture, and low-speed autonomy. It is less suitable for arbitrary urban driving where the vehicle must choose new routes dynamically.

Implementation Notes ​

  • Store raw scans or reprocessable features so route maps can be regenerated after algorithm updates.
  • Keep teach and repeat calibration metadata; radar mounting and timing changes affect localization.
  • Build explicit route-version management and operational rollback.
  • Use local route-relative metrics in addition to global ATE: lateral error, heading error, along-track error, intervention rate, and path-tracking RMSE.
  • Test route segments with decreasing structure, not only visually distinctive areas.
  • Keep radar odometry, route localization, and path-tracking health separate in telemetry.
  • Treat radar teach-repeat as a route autonomy layer inside a broader safety system, not as a complete perception stack.
  • Re-check CFEAR-TR code status before implementation planning. The arXiv record says a C++ implementation is available, but this pass did not find a clearly linked TR-specific public repository; older CFEAR radar-odometry repositories are adjacent but not necessarily the same release.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim