A two-class CPC is a signal to read the claim narrowly, and GM's grant US10678253B2, "Control systems, control methods and controllers for an autonomous vehicle" (issued June 9, 2020), carries just two: G05D 1/0223 (autonomous control responsive to the environment) and G01C 21/20 (navigation). On its face that keeps the claim on the control layer — turning a plan and a position into vehicle motion. The actual independent claims, however, are considerably more specific than the title suggests, and they reach back into perception in a way the classification alone does not advertise.

Read independent claim 1 and the architecture comes into focus. The method runs on a "high-level controller that executes a two-stage neural network." The first stage produces a feature map; the second stage produces a perception map derived from that feature map. The claim then pins down exactly what the feature map is: "a concatenation of a range presence map derived from at least one of a radar system and a lidar system, a current vision-based feature map derived from images captured by at least one camera at the particular instant of time and a previous vision-based feature map derived from images captured by the at least one camera from a previous time instant." That single limitation is doing a lot of work. It is an early-fusion claim — radar/lidar occupancy concatenated with two camera frames, current and prior — and the inclusion of a previous-frame feature is what injects temporal context before any object is ever drawn.

“Systems and methods are provided for controlling an autonomous vehicle (AV). A map generator module processes sensor data to generate a world representation of a particular driving scenario (PDS).”— U.S. Patent No. 10,678,253 source

The second stage converts that machine-readable feature map into a human-readable perception map, and the claim enumerates its contents rather than leaving them abstract: "bounding box location, orientation, and velocity of each detected object," "an image segmentation of freespace," "road feature locations/types," and "stixels that approximate object boundaries." Stixels — thin vertical segments standing in for obstacle boundaries — are a specific, citable construct, not boilerplate. By naming velocity at the bounding-box level and freespace segmentation in the same breath, the claim couples detection, motion, and drivable-area extraction into one map handed downstream.

Where the claim earns its breadth is the control half. The high-level controller "decompos[es] the autonomous driving task into a sequence of sub-tasks," then selects, from a library of "sensorimotor primitive modules" (SPMs), "a particular combination" to enable for the current scenario. Each SPM "directly maps information from the feature map to one or more control actions" yielding "a particular vehicle trajectory and speed profile." This is the load-bearing idea: rather than one monolithic planner, the controller composes a maneuver from reusable primitives selected on the basis of the feature map, and crucially the SPMs key off the feature map directly — not the cleaned-up perception map — so the control primitives consume raw fused features. The selected trajectory-and-speed profile then becomes control signals, and a separate "low-level controller" turns those into actuator commands. Claim 3 names those actuators concretely: "a steering controller, a brake system, and a throttle system." The two-controller split — high-level composition, low-level actuation — is the structural backbone repeated across the system claim (claim 4) and the controller claim (claim 7).

So the early read that this is "narrow because two CPCs" was half right and half wrong. It is narrow in the sense that examination forced a long string of architectural limitations into claim 1 — a granted B2 with a claim this detailed did not survive by being broad. But it is not a thin control-law claim; it is a perception-to-control pipeline claim whose novelty sits in the specific feature-map concatenation and the sensorimotor-primitive composition step. The G05D 1/0223 / G01C 21/20 pairing captures the navigation-to-control surface but understates how much perception structure the claim actually recites.

On scope and dependents, the moat is layered. Claims 2, 5, and 8 reinforce that each primitive produces a trajectory-and-speed profile for "a specific driving maneuver," tying the primitives to discrete maneuvers. Claims 3, 6, and 9 lock the actuator triplet to the low-level controller across all three statutory classes (method, system, controller), so a competitor cannot dodge by recharacterizing the same pipeline as a "controller" rather than a "method." A would-be designer-around has to avoid both the concatenated multi-modal/temporal feature map and the select-and-execute-primitives composition to land outside this grant — and avoiding both while still building a learned AV motion stack is a tight squeeze.

It is worth being precise about why the claim reaches into perception while wearing a control classification. G05D 1/0223 covers control responsive to the environment and G01C 21/20 covers navigation, so a classifier reasonably tags the output side of the pipeline. But the limitations recite the input side too — the two-stage network, the multi-modal/temporal feature concatenation, the enumerated perception-map contents — because the claim's inventive step is the end-to-end coupling: features in, composed maneuver out, actuators last. Splitting that into a separate perception class would have understated the integration that is the whole point. The temporal element deserves a second look as well: by demanding both a current and a previous vision-based feature map, claim 1 builds short-horizon motion cues into the representation before object detection runs, which is what lets the downstream velocity-bearing bounding boxes and stixels carry consistent dynamics. That is a deliberate architectural choice, not incidental wording, and it is the kind of limitation an examiner would have leaned on to allow the case.

Strategically, GM filing steadily on control-layer IP in 2020 — alongside its monitoring and rule-learning grants from the same year — fits a company protecting the full autonomous stack rather than only the perception headline. The verdict: enabling, examined, and more architecturally specific than its title or its two CPC tags imply. Read the substance in the feature-map concatenation and the sensorimotor-primitive selection; treat the generic abstract language as scaffolding, not as the invention.