perception, hardware, eval
Adding a second sensor does not delete uncertainty. It moves the uncertainty into a referee.
Send two scouts to watch the same opponent. One says they will press high. The other says they will sit deep. You now have more reports and a new job: decide whom to believe, with no third scout to break the tie.
Combining sensor types — often called fusion — has that shape. Most of the time the feeds agree and confidence rises. The hard cases are the minority of moments when they do not, and those moments cluster around the scenes that matter.
One says they will press high.
A textbook conflict: a ranging sensor reports a stationary object in the lane. A camera sees an overhead sign. Trust the ranging sensor and the vehicle may brake hard on an empty highway. Trust the camera and you may have ignored a stopped vehicle.
Every tie-break rule has a cost. Always prefer the more cautious feed and you get extra braking for ghosts. Weight the feeds by learned confidence and you have added another component that can be wrong. Require agreement before acting and you slow the system when time is short.
Fusion can still be a good design. The point is narrower: more sensors are more information plus an arbitration problem. The arbitration logic becomes a new place for errors to live.