Availability-Aware Sensor Fusion

What It Is

Availability-aware Sensor Fusion, or ASF, is a 4D radar, LiDAR, and camera fusion method.
The full paper title is "Availability-aware Sensor Fusion via Unified Canonical Space for 4D Radar, LiDAR, and Camera."
It was released as a 2025 arXiv paper by KAIST researchers.
The method targets object detection when one or more sensors are degraded or unavailable.
It is explicitly designed around sensor availability instead of assuming all modalities are always valid.
The main benchmark in the paper is K-Radar.

Deeply coupled fusion can fail when one sensor disappears or degrades.
Sensor-wise cross-attention can avoid coupling but is often expensive and hard to align.
ASF projects camera, LiDAR, and 4D radar features into a unified canonical space.
Unified Canonical Projection, or UCP, makes features from different sensors more consistent before fusion.
Cross-Attention across Sensors Along Patches, or CASAP, fuses available sensor features patch by patch.
Sensor independence is preserved so unavailable sensors do not poison the fused representation.
The canonical feature space lets cross-attention operate with lower cost than all-pairs sensor fusion.

Inputs are camera images, LiDAR point clouds, and 4D radar point or tensor features.
The method can run with degraded or missing sensor modalities in the evaluated conditions.
Intermediate outputs are canonical-space sensor features and patch-level attention features.
Final outputs are 3D object detections with BEV and 3D AP metrics.
The paper reports AP BEV and AP 3D at IoU 0.5.
The method assumes a calibrated multi-sensor rig and a dataset with all three modalities.
It is not a sensor fault diagnosis model, although it is designed to tolerate faults.

Each sensor has its own feature encoder before canonical projection.
UCP maps each modality into a shared feature representation.
CASAP performs cross-attention among sensors along spatial patches.
The design avoids direct feature coupling that would make one feature tensor depend irreversibly on a failed sensor.
A detection head predicts objects from the availability-aware fused representation.
The paper compares ASF to deeply coupled fusion and sensor-wise cross-attention baselines.
Experiments include adverse weather and sensor degradation or failure conditions.

Evaluation is performed on the K-Radar dataset.
K-Radar provides camera, LiDAR, and 4D radar data in multiple weather conditions.
The paper reports ASF reaching 87.2 AP BEV and 73.6 AP 3D at IoU 0.5.
Reported improvements are 9.7% in AP BEV and 20.1% in AP 3D over prior fusion methods in the paper's setup.
The key evaluation question is whether the detector retains performance when sensor availability changes.
The arXiv page states that code will be available through the KAIST AVELab K-Radar repository.
Production replication should verify the exact release status and configuration before comparison.

Covers camera, LiDAR, and 4D radar in one fusion design.
Radar inclusion makes it more relevant to adverse weather than camera-LiDAR-only methods.
The method explicitly acknowledges missing and degraded sensors.
Patch-level attention offers a practical tradeoff between robustness and compute.
Canonical projection makes sensor feature alignment an architectural primitive.
The K-Radar evaluation includes weather variation that is closer to operational degradation than clean nuScenes only.

Results depend heavily on K-Radar sensor layout, labels, and weather distribution.
Canonical features may still encode dataset-specific modality biases.
A missing sensor is different from a subtly wrong sensor; the latter can be harder to detect.
Calibration drift and timestamp skew can still corrupt canonical projection.
4D radar quality varies widely across hardware vendors and mounting locations.
The method does not provide formal uncertainty bounds or safety guarantees.
If all available sensors are weak in the same spatial patch, CASAP has no reliable evidence to recover.

ASF is a strong conceptual fit for airport apron AVs because radar is useful in fog, rain, spray, darkness, and glare.
Availability-aware fusion maps naturally to ODD policies that degrade speed or autonomy when sensors are impaired.
It could support sensor-suite designs that keep radar as a primary resilience modality rather than a late-stage add-on.
Airside transfer requires data with aircraft, ground crew, tugs, cones, baggage carts, jet bridges, and reflective clutter.
Apron weather and lighting should be evaluated as real sensor data, not only synthetic corruption.
ASF should be paired with explicit sensor-health monitoring and planner-side capability limits.

Check whether the exact ASF code and configs are released before planning reproduction.
If porting to an airside stack, keep modality masks explicit and logged.
Benchmark single-sensor, two-sensor, and all-sensor cases separately.
Add calibration and timestamp perturbations to the evaluation because canonical projection depends on geometry.
Verify radar preprocessing for the target hardware, including Doppler channels and elevation resolution.
Use degraded-input validation from MSC-Bench, MultiCorrupt, and S2R-Bench as complementary tests.