ADR 0004: hybrid estimation with VIO, gate PnP, and a drag-aware model

Proposed · 2026-10-23 · racing-drone/docs/adr/0004-gate-relative-hybrid-estimation.md

Status: Proposed Date: 2026-10-23 Related: requirements EST-03..08; cortex ADR 0012

Context

MonoRace won with gate PnP and a model-based EKF, and no VIO. It crashed after more than 25 m without valid images. Cortex already has an MSCKF VIO, but it is over-confident (#212) and five times too slow for 90 Hz. Accelerometers saturate at around 7 g.

Decision

• Two estimator implementations behind the Estimator slot:
  • GateEkf: MonoRace-style. 16 states, drag-model prediction, delay compensation, and IMU-saturation switching. During saturation, prediction uses the model’s specific force instead of the accelerometer, and the position and attitude covariance is inflated so that vision updates dominate. VioGateEkf uses the same switching.
  • VioGateEkf: the cortex MSCKF fused with gate PnP and the same model and saturation handling.
• Both express their state in the course frame, anchored to the gate map.
• Promotion criterion. VioGateEkf is the primary estimator only if it beats GateEkf as follows:
  • Comparison set: the same baseline (B2 vs. B2m, identical controller and parameters) on the test-plan sensor-degradation group. That is camera corruption at 25 % and 50 % of frames, and 0.5 s and 1 s blackouts. Run 200 paired seeds per scenario in L3 (the same seed for both estimators).
  • Required gain: two-lap success rate at least 5 percentage points higher, or true gate-crossing error (|bias| + 3σ) at least 20 % lower, pooled over the set, with neither metric worse.
  • No regression: on the nominal group, success rate no more than 1 percentage point lower, and both estimators still meet EST-04 / EST-05.
  • EST-04 / EST-05 during blackouts: they apply at the first gate crossed after the blackout ends, not during it. NEES consistency (EST-05) must hold throughout, including the blackout. That is the point of VIO: a covariance that grows honestly.

Consequences

• Phase 3 does not depend on fixing VIO. B2m can fly on GateEkf alone.
• Cortex VIO work (NEES, latency, rolling shutter) is prioritized by a mission metric rather than by EuRoC ATE.

Current state: proposed.