Vision-only versus mapped is a cost bet, but diversity-versus-single-method is a safety bet — and Nvidia placed it. US11520345B2, granted in December 2022, covers "Path perception diversity and redundancy in autonomous machine applications."
The thesis is that a single way of perceiving the road is a single point of failure. Classified under G05D 1/0219 (route planning), G06V 20/588 (lane/road perception) and G06N 3/08 (neural networks), the patent describes perceiving the drivable path through diverse, independent methods — so that when one method is fooled by glare, faded markings or an unusual scene, another still sees the road.
Here is the engineering insight worth keeping. Redundancy only helps if the redundant systems fail differently. Two copies of the same perception model fail on the same inputs — they share a blind spot. Diversity is the harder, more valuable property: methods built on different principles fail on different scenes, so the system as a whole has fewer common blind spots.
The honest cost is compute and complexity. Running multiple diverse perception methods burns silicon and demands logic to fuse or arbitrate when they disagree. For Nvidia, whose business is selling that silicon, the cost is conveniently aligned with the product — a path-perception architecture that needs more compute is, for the arms dealer of autonomy, a feature.
That self-interest does not make the safety argument wrong. Aviation has used dissimilar redundancy for decades for exactly this reason, and the patent imports that discipline into driving perception. It is a more rigorous answer than the single-model end-to-end pitch that gets the demos.
For readers in the sensor-wars trenches, this patent reframes the debate. The fight is not only camera-versus-lidar; it is also single-method-versus-diverse-methods. A stack that perceives the road two genuinely different ways is more defensible than one that perceives it once, however good that one way is.