The independent claim is where the value lives, and in GM's grant US10558217B2, "Method and apparatus for monitoring of an autonomous vehicle" (issued February 11, 2020), the value is a supervision layer rather than a driving capability. The claim is not about how the car perceives the road; it is about a separate function that judges whether the perception-and-control pipeline is performing within bounds. Read it as a watcher built to catch the watcher failing.
That distinction matters because the CPC tells the same story. The grant is classified under G05D 1/0088 (control of autonomous land vehicles) alongside G06K 9/00791 and G06K 9/00201 — road-scene and image recognition classes. A patent that combines an autonomous-control class with image-recognition classes and frames the invention as monitoring is describing a self-diagnostic loop, not the primary driving function. The defensible novelty is the monitoring construct, and a teardown should not overstate it as a claim on autonomy itself.
Why file this in 2020? Because by then the industry had learned that the hard part of an automated stack is not the happy path but the failure detection. A perception model that silently degrades — fog, sensor occlusion, an out-of-distribution scene — is more dangerous than one that fails loudly. GM's claim positions the company on the supervision problem, which is exactly the layer regulators and safety cases scrutinize. That is a strategic place to hold IP.
The dependent claims are the moat here. They narrow the monitoring to specific signals and thresholds, and a dependent that ties a fault flag to a defined control response is the kind of limitation that is genuinely defensible because it reads on a concrete engineering choice. The breadth lives in the independent claim; the durability lives in the dependents. Strip the abstract's broad language and what GM actually obtained is a position on onboard self-supervision — a quiet but well-chosen corner of the autonomous-driving record.