The unglamorous step is where the claim lives. Torc Robotics' grant US12651462B1, "Systems and methods for point cloud annotation using kinematic models" (issued June 9, 2026; inventors David John Thompson, Robert Charles Kriener, Haseeb Chaudhry), claims a method for labeling LiDAR point clouds — the training data — by applying kinematic models of how objects move. Not a perception model; the annotation that trains one.
The mechanism, plainly: to train a perception network you need labeled data — point clouds where each cluster is tagged as a car, a pedestrian, a cyclist. Labeling LiDAR by hand is brutally expensive. Torc's method uses kinematic models — physics-based descriptions of how objects move — to propagate and infer labels automatically across frames. The CPC reflects it: G06V 20/56 for street-scene recognition, G01S 17/89 for the 3D LiDAR source, G06T 5/50 for image combination, G06V 20/70 for labeling-related processing.
Why patent the labeling and not the model? Because in autonomy, the data pipeline is a real moat and an under-watched one. Perception architectures turn over fast and are heavily published; the proprietary, defensible advantage often lives in how cheaply and accurately a company can manufacture training labels at fleet scale. A grant on kinematic-model annotation fences a piece of that pipeline — the part that determines how fast and how cheaply the perception stack improves.
For a strategy read, this is the kind of claim that signals where a company thinks its durable edge is. Torc filing on annotation rather than only on perception suggests it values the data-factory advantage. It also rhymes with a pattern across the sector: the boring infrastructure steps — calibration, annotation, synthetic data — are quietly accumulating IP, because they're the parts that don't get obsoleted every model cycle. The claim ages more slowly than a perception claim would.
Caveats, honestly. An annotation-method claim is bounded by the specific use of kinematic models it recites — a competitor labeling by a different automated method may sit outside, and you confirm the boundary by reading claim 1. The B1 kind code means no pre-grant publication, so the prosecution history is the place to see how the examiner shaped it. And owning an annotation method says nothing about data volume or label quality in practice.
For the control-and-autonomy beat: don't skip the plumbing patents. Torc's grant fences the data-labeling layer that every perception model depends on — slower-aging and more strategic than it looks. Pull it on PatentBear, read what claim 1 says about the kinematic-model annotation, and file it under data-pipeline IP, the quiet category where a lot of autonomy's real moats are being built.