The probe's true track (green) is known only to this simulation. Its onboard estimate (blue, wrapped in a wireframe uncertainty ellipsoid) is all the spacecraft itself has to steer by. Every step it dead-reckons forward with its inertial sensors — velocity plus a slowly wandering gyro/accelerometer bias — so the estimate's error grows like a random walk. Whenever three or more surface landmarks are simultaneously inside the camera's field of view and lit well enough to detect, an optical fix pulls the estimate back toward the truth and shrinks the ellipsoid; more visible landmarks mean a tighter, faster correction.
predict: σ² += (driftRate · dt)²
update: k = visibleCount / 25 (capped)
error *= (1 − k)
σ² = max(measNoise² / visibleCount, σ²·(1−k)²)
- IMU drift rate — how fast the inertial dead-reckoning bias wanders; higher rates blow the ellipsoid up faster between fixes.
- Sun elevation — moves the terminator across the surface; near the extremes most landmarks fall into shadow or glare and the visible count collapses.
- Active landmarks — how many candidate surface features the mapping pass has catalogued; a sparse landmark set means fewer fixes qualify and gaps between corrections widen.
- Camera updates OFF — disables the optical fix entirely, showing pure inertial dead reckoning drift without bound, the failure mode landmark-based nav exists to prevent.
Real missions (OSIRIS-REx, Hayabusa2, Rosetta) run exactly this predict/update cycle — star trackers and IMUs propagate attitude and position between opportunities, and optical landmark or terrain-relative fixes reel the covariance back in whenever geometry and lighting allow.