Each sensor reports a live confidence score (0–100%) based on whichever hazard currently affects it. Dust and glare specifically blind the camera; rain adds LIDAR ranging noise and confuses the ultrasonic echo off wet, reflective puddles. A camera-only robot's overall perception reliability is its camera confidence — when that hits zero, the robot is effectively driving blind. A sensor-fusion robot instead combines all three:
reliability = 1 − (1−camConf)·(1−lidarConf)·(1−ultraConf)
This treats perception as failing only if every sensor fails at once — so as long as one sensor stays clean, the fused estimate stays high and steering noise stays low. The robot's actual obstacle-sensing is corrupted with random error proportional to (1 − reliability), so low reliability directly causes wandering and collisions, not just a cosmetic number drop.
- Dust cloud — obscures the camera image; mildly scatters LIDAR returns.
- Glare — direct sun saturates the camera sensor; LIDAR and ultrasonic unaffected.
- Rain / mist — adds LIDAR range noise and confuses ultrasonic off wet surfaces; camera only mildly softened.
- Replay sequence — runs dust → glare → rain back to back, so you can switch modes and compare the same hazard timeline.