Experiment: evidence grounded radar depth refinement under degraded visibility

Problem

Radar remains geometrically useful in smoke, fog, darkness, and visual occlusion, but angular resolution is limited. A generative prior can recover structure, but it can also hallucinate plausible geometry that is inconsistent with measured range.

GRADE, submitted 2026-09-09 and listed to appear at ACM MobiCom 2026, grounds latent diffusion refinement in single frame 4D radar geometry and uses residual camera cues only when available.

Paper: https://arxiv.org/abs/2609.10756
Project: https://phi-lab-rice.github.io/GRADE

Research evidence

The authors report about 95K synchronized frames across 12 buildings with real smoke. The abstract reports metric depth MAE of 0.303 m in clear scenes and 0.313 m under smoke. The paper and project materials also show that radar geometry must remain the anchor; generative refinement without adequate measurement conditioning can improve visual structure while worsening metric placement.

These are external paper results, not RuView results.

Current architecture

RuView already supports mmWave, LiDAR integration work, RF spatial memory, world model experiments, and an evidence hierarchy. It does not yet have a benchmarked contract for generative spatial refinement that prevents priors from overriding measured radar geometry.

Proposed improvement

Build an experimental radar_depth_refine path with three strictly separated stages:

1. RadarMetricDepth: deterministic or learned coarse metric depth derived only from measured radar.
2. StructuralPrior: optional generative refinement that receives the coarse metric estimate as an immutable condition.
3. EvidenceGate: rejects or clips refined geometry when residuals exceed range, uncertainty, or observability bounds.

The output must retain both the measured coarse depth and the refined hypothesis. Refined confidence may not exceed source evidence confidence unless an independent sensor supplies supporting evidence.

Expected measurable improvement

On held out buildings and visibility conditions, promote only if the refined path achieves both:

• at least 15% lower depth MAE than the existing radar only reconstruction baseline
• no more than 2% increase in catastrophic geometric outliers beyond the configured range tolerance

Under camera blackout or heavy degradation, the fused path must converge toward the radar anchored output instead of preserving visual hallucinations.

Dependencies

Prefer existing RuView radar, LiDAR, model runtime, and provenance components. Do not add the GRADE model or a diffusion dependency to production until license, model weights, memory, latency, and cross platform viability are verified.

Security and privacy

• raw radar cubes remain P0 and local by default
• external model weights require integrity hashes and supply chain review
• malformed radar tensors, NaN, infinity, oversized dimensions, timestamp disorder, and sensor frame mismatches fail closed
• generative output cannot create an identity or health inference absent separately authorized evidence
• camera cues are optional and must preserve their own privacy classification
• adversarial RF inputs must not cause unbounded diffusion or allocation work

Testing

• unit tests for metric residual and confidence monotonicity gates
• property tests that refinement cannot move a pixel or point outside configured uncertainty without independent support
• synthetic smoke is not sufficient for promotion
• held out building split
• real smoke or fog test where available
• camera blackout and corrupted camera inputs
• radar dropout and corrupted calibration
• x86 GPU benchmark first; ARM64 edge inference only if a compact model is demonstrated

Benchmark plan

Compare:

1. current radar depth baseline
2. radar only learned baseline
3. ungrounded generative refinement control
4. measurement grounded refinement
5. grounded refinement plus optional camera cues

Measure MAE, p90 absolute error, Chamfer distance, catastrophic outlier rate, ECE, p50 and p95 latency, VRAM, RAM, model size, startup, power where measurable, and degradation across buildings.

Use at least three runs for runtime metrics and report dataset split hashes.

Compatibility and rollout

Research feature flag only. No existing output or safety claim changes. Keep coarse radar geometry and current main as rollback.

ADR requirement

Create a focused ADR before code promotion covering evidence hierarchy, confidence monotonicity, generative prior boundaries, tensor contracts, model integrity, degraded sensor behavior, and rollback.

Definition of done

• GRADE evaluation protocol independently reproduced or discrepancy documented
• license and model artifact provenance verified
• five arm RuView benchmark completed on a building disjoint split
• source geometry remains queryable beside refined geometry
• no confidence laundering
• security review complete
• production path remains disabled until real hardware evidence passes

Production classification

Research only initially. The dataset and real smoke evidence are unusually strong, but diffusion latency, model size, licensing, and RuView hardware transfer must be measured before promotion.

No source files or entry points are named. Start by tracing the existing radar, LiDAR, model runtime, and provenance components, then create the required ADR before implementation. Done means the five-arm benchmark, security and provenance checks, queryable coarse and refined geometry, and disabled production rollout are complete.