The expensive part of building a self-driving system is rarely the sensor. It is the labeled data that teaches the perception model what it is looking at, and the human hours that labeling consumes. So when a company is issued an enforceable claim on a way to automate that labeling and feed it straight into the part of the system that decides how the vehicle moves, that is a piece of the pipeline worth reading closely. On June 2, 2026, Aurora Operations, Inc. received exactly that in US12643575B2, a method for automatic annotation of object trajectories in multiple dimensions.

The claim describes a system that takes sensor data — the patent specifies sequential point clouds of lidar data — generates an initial trajectory for a surrounding object, fixes a value for that object's size, refines the trajectory against the fixed size, and produces a multi-dimensional label. The label is not the end of it. The same record carries that label forward into vehicle control, closing the loop from raw perception to the driving decision.

A motion plan for controlling the AV can be generated based on the multi-dimensional label.— Automatic annotation of object trajectories in multiple dimensions, US12643575B2

The classification places the grant across the B60W 60/0027 family (vehicle operation in response to detected objects) and the G05D 1/02 path-control classes, with G06N 3/08 and G06N 20/00 marking the machine-learning method underneath. For a reader mapping where autonomy content is being locked up, this is a claim on the annotation-to-planning bridge — the connective tissue between what the truck sees and what it does. Why that bridge matters commercially is a question of where the labor goes. Autonomy programs spend enormous effort building the ground-truth datasets that supervise their perception models; every frame a human has to draw a box around is cost, and the size of that cost is one of the quieter line items separating the companies that can scale from the ones that stall. A claimed method that generates refined, size-consistent, multi-dimensional labels from sequential lidar point clouds — and that the grant explicitly carries into a motion plan — is coverage on a piece of exactly that cost structure. The inventor list on the grant, which includes researchers associated with Aurora's perception group, is consistent with a core-pipeline filing rather than a peripheral one.

The same week: lidar, prediction, tracking

The annotation grant did not arrive alone. The same window of Aurora issuances reaches into the layers on either side of it. US12649490B2 covers generating behavioral predictions for surrounding actors that react to the autonomous vehicle's own planned movement — a prediction model conditioned on a candidate motion plan, which is the prediction side of the same perception-planning loop. US12649476B2 claims improved tracking of articulated vehicles, modeling a tractor and its trailer as linked portions — a notably trucking-specific problem for a company whose commercial product is autonomous freight.

Further up the pipeline, Aurora's lidar hardware shows up in the record the same period. US12656601B2 claims a method for optimizing the scanning of coherent lidar using two scanners rotated at different angular velocities, and US12638559B2 covers an optical coupler for a lidar sensor — both in the G01S 17/931 class that dominates Aurora's portfolio. These are the frequency-modulated continuous-wave (FMCW) lidar internals that Aurora has built much of its sensing identity around. That the same week's grants reach from the lidar's internal optics all the way to the annotation method and the motion plan is worth stating plainly, because vertical coverage of that kind is uncommon. Many autonomy companies buy their sensors and file only on software, or build sensors and license the stack; the record here shows Aurora holding issued claims at both ends and in the middle. The presence of the evaluation platform (US12646363B2) and the distributed-model method (US12645929B2) in the same window rounds out the picture with the operational tooling — how the system is tested against logged scenarios and how its models are run across compute — that a fleet program needs but rarely advertises.

What the spread of classes shows

Read together, the week sorts into a recognizable shape: sensor optics (the lidar grants), perception and data labeling (the annotation method), prediction (the reactive-behavior grant), tracking (the articulated-vehicle method), and the supporting infrastructure that makes a fleet program legible — US12646363B2, a lightweight platform for evaluating an autonomous-vehicle control system against logged scenarios, and US12645929B2, a method for segmenting machine-learned models across distributed compute. A single week of grants touching the sensor, the labels, the prediction, the tracker, and the test harness is the data point.

That spread is consistent with the broader record. Aurora's issued footprint in the patent record runs to roughly three hundred grants, with the largest single cluster — about ninety-nine records — in the G01S 17/931 lidar class, and substantial counts in the G05D 1/0088 autonomous-control and B60W 60/001 operation classes. The annotation patent adds to the machine-learning-method side of that portfolio, the G06N classes where the labeling and prediction work lives. It also helps to size the new grant against the boring incumbent activity in the same window. The week's broader robotics-and-autonomy grant flow was led on volume by a chip supplier whose patents recur as enabling components across many companies' systems; against that backdrop, a single operator filing a vertically integrated set of claims — sensor optics, perception, prediction, tracking, and the planning bridge between them — is a distinct pattern, and one specific to a company that builds and runs its own stack rather than supplying parts of others'. The articulated-vehicle tracking grant (US12649476B2) is a small but pointed example: it is a claim shaped by the trucking use case, where a tractor and trailer move as linked bodies, not by the passenger-car problems that dominate the rest of the field's filings.

None of this speaks to how the claims will be enforced, or whether they will be. What the June 2 record establishes is narrower and firmer: as of that date, Aurora holds issued coverage that runs the length of its autonomy pipeline — from the coherent-lidar scanner to the multi-dimensional label to the motion plan — documented in patent numbers with the company on the assignee line.