Most discussion of self-driving sensors in bad weather focuses on the perception model — can the neural network still recognize a pedestrian through rain spray? GM Cruise's latest grant addresses an earlier, quieter question: can the LiDAR itself tell that the weather has changed, before any object-recognition step runs? U.S. Patent No. 12,656,500, issued June 16, 2026 and assigned to GM Cruise Holdings LLC, claims a method for detecting weather conditions "at a LiDAR sensor level" by treating the point cloud as a statistical distribution and watching that distribution drift. The named inventors, Daniel Flores Tapia and Rebekah Brandt, frame it as a sensor-level integrity check rather than a perception feature.
The mechanism is concrete and refreshingly old-school in its statistics. The system first builds a reference probability mass function — a histogram, essentially — over at least one field of a point cloud captured from a clean reference scene. As the vehicle drives, it computes the same probability mass function for the current scene. It then measures how far apart the two distributions have moved using a Kullback-Leibler divergence calculation, a standard information-theoretic measure of how one probability distribution diverges from another. When that divergence crosses a defined threshold for the field in question, the method flags an environmental change in the current scene response. Rain, fog, snow, and spray all alter the population of returns a LiDAR sees — more short-range scatter, dropped long-range points, shifted intensity — so a distribution that has wandered from its clean-weather reference is a usable proxy for "conditions just degraded."
"Disclosed are systems and methods for detecting weather conditions at a LiDAR sensor level."— U.S. Patent No. 12656500, source
What the independent claim actually fences off
Read the limitations in order and the scope becomes clear. Calculate a reference PMF of at least one field of a point cloud generated from reference scene responses. Calculate a current PMF for that same field from a current scene response. Determine a statistical difference between the two using a KL-divergence calculation. And, responsive to that difference satisfying a threshold, flag an environmental change. The claim's distinctiveness is the chain: it is not merely "detect weather with LiDAR," it is the specific pairing of a per-field point-cloud distribution with a KL-divergence threshold test as the detection rule. That specificity cuts both ways for scope — it is harder to argue infringement against a system that uses a learned classifier or a different statistical distance, but it is also a clean, defensible claim against anyone who reaches for the same information-theoretic shortcut.
The CPC classifications corroborate the framing. The grant sits in G01S 17/95 (LiDAR for meteorological or atmospheric measurement), G01S 7/4802 and G01S 7/4861 (LiDAR signal processing and receiver circuitry), and G01S 17/10 (pulsed ranging). This is classed as a sensing-and-signal-processing invention, not a perception or planning one — which matches the claim's insistence on operating "at a LiDAR sensor level," upstream of the object-detection stack. The point is to make the sensor self-aware about its own measurement quality.
Why a sensor-level weather check matters — and what it doesn't tell us
For the mapped-LiDAR autonomy camp, weather is the edge case that erodes the whole value proposition. A robotaxi that has to disengage or pull over the moment it rains is not a transportation service; it is a fair-weather demo. A cheap, fast, model-free signal that conditions have degraded is exactly the kind of fallback trigger an autonomy system wants, because it can drive a minimal-risk decision — slow down, widen following distance, hand back control, or down-weight LiDAR in favor of other sensors — without waiting for the perception network to fail in some harder-to-detect way. Placing that check at the sensor, using a deterministic statistical test rather than a learned one, also makes it cheaper to validate and easier to reason about for safety cases. That is a real engineering virtue when regulators want to know exactly when and why the system changes its behavior.
There is an unavoidable corporate footnote here. GM Cruise wound down its robotaxi ambitions, and the assignee on this grant, GM Cruise Holdings LLC, is the legacy entity; the patent reflects work done while the program was active and prosecuted to grant on the June 16, 2026 issue date. That makes the document more interesting as IP than as product signal. A granted, well-bounded weather-detection method does not expire because a robotaxi program did — it becomes an asset that can be asserted, licensed, or folded into GM's broader driver-assistance portfolio, where the same LiDAR-integrity logic is directly useful. For anyone mapping the autonomy patent landscape, the lesson is to watch the assignment and the claim, not the press cycle: the program may be gone, but the fenced-off method is now on the books.
The takeaway for a freedom-to-operate read is the test, not the title. "Detecting adverse weather conditions at a lidar sensor level" is broad-sounding; the granted claim is a precise recipe — reference PMF, current PMF, KL divergence, threshold, flag. Build weather detection some other way and you are likely clear. Reach for that exact statistical shortcut on a LiDAR point cloud and this is the grant a diligence search will surface.