Skip to content

SPARSE_SCALE_CLEARANCE_SIDECAR_R0_RESULT.md

Latest commit

 

History

History
83 lines (64 loc) · 3.44 KB

SPARSE_SCALE_CLEARANCE_SIDECAR_R0_RESULT.md

File metadata and controls

83 lines (64 loc) · 3.44 KB
Raw

Sparse-scale clearance sidecar R0

Date: 2026-08-03

Terminals:

  • SPARSE_SCALE_CLEARANCE_SIDECAR_REPLAY_SUPPORTED_CONSUMED_PROXY
  • CLOCK_DOMAIN_BINDING_REQUIRED_AND_VERIFIED
  • REAL_TOF_REGISTRATION_NOT_EVALUATED

Decision

Retain the following conditional candidate for hardware integration:

calibrated external RGB
  -> DA V2 Small Metric Hypersim 392x518
  -> left / center / right raw clearance
  + timestamped sparse metric-scale anchors
  -> missing, future, stale, or empty input => UNKNOWN
  -> scaled three-band clearance sidecar (no alert)

This is the quality/cost candidate, not a claim that the smaller observer is more accurate than Metric3D. Metric3D remains the stronger standalone observer and current teacher/reference. The retained hypothesis is that a faster, smaller observer plus a cheap recurring metric anchor can satisfy the frozen task gates at lower deployment cost.

End-to-end consumed replay

The new class-free sidecar recomputed all 120 consumed TUM RGB frames through the 392x518 PyTorch observer and clearance geometry. It did not reuse stored candidate clearance fields. One fixed anchor was materialized at frame 9 of each 30-frame sequence from the already-consumed registered sensor-depth proxy; frames 10-29 were evaluated.

Measure Result
eligible / paired-valid frames 80 / 78 (97.5%)
clearance MAE 0.098145 m
collision agreement 93.7729%
false-clear rate 4.9451%
temporal clearance-delta MAE 0.085803 m
frozen task gates 5 / 5
host depth median 33.0685 ms
host geometry + scale median 11.3147 ms

The host timings are PyTorch replay diagnostics and cannot be added to or substituted for the separately measured HTP model timing. They do not establish camera-to-clearance latency, energy, thermals, or sustained frame rate.

Clock and failure behavior

The first local integration attempt bound anchor timestamps to the absolute image clock while the manifest used sequence-relative timestamps. It failed closed: 119 frames were UNKNOWN_NO_METRIC_SCALE_ANCHOR and one was UNKNOWN_RAW_CLEARANCE; no incorrectly scaled output was emitted.

The materializer now binds anchor time to the manifest authority. The corrected replay produced 36 pre-anchor UNKNOWN_NO_METRIC_SCALE_ANCHOR, 83 VALID, and one UNKNOWN_RAW_CLEARANCE, with no VALID row before the anchor frame. Monotonic ordering is enforced, and unit fault injection verifies that an expired anchor becomes UNKNOWN_STALE_METRIC_SCALE_ANCHOR.

The replay used an explicit 5000 ms maximum anchor age only to keep the fixed single-prefix proxy available over each short sequence. That value was not selected as a deployment TTL and must not be inherited by real hardware.

Claim ceiling and next step

The anchor is a registered sensor-derived proxy, not real multi-zone ToF. The observer ran on the host, not through an end-to-end HTP camera pipeline. This result therefore supports interface causality and consumed task quality only; it does not validate final-camera optics, RGB-ToF spatial registration, anchor availability, synchronization jitter, device cost, alerts, safety, production, research-mainline promotion, or the default App.

The next valid experiment is a real multi-zone ToF/RGB registration adapter and device replay with a prospectively fixed expiry policy. Another depth-model search is not authorized by this result.

Machine-readable result: SPARSE_SCALE_CLEARANCE_SIDECAR_R0_RESULT.json.