A granted patent is enforceable coverage, so when two issue on the same date to one autonomous-driving company, the question for a business reader is where in the stack the company is fencing. For Five AI — a UK autonomous-driving software firm — the two patents issued under the March 17, 2026 grant date point not at the car's perception or control, but at the apparatus used to test the software that drives it. That emphasis is the defining feature of the company's portfolio, and it is worth mapping.

The clearer of the two is US12576864B2, "Tools for performance testing and/or training autonomous vehicle planners." It covers running a target planner across incrementing planning steps in a scenario, then evaluating it by computing a reference plan for a given instant and scoring how the planner's plan compares. Its classification — B60W 50/06 (conducting tests on vehicle subsystems) with B60W 60/0013 and B60W 60/0015 (autonomous trajectory planning) — places it squarely in the testing-of-planners corner. The companion grant, US12576886B2 ("Motion prediction"), covers predicting an agent's behavior by projecting its measured velocity onto a reference path to forecast where it will go, and generating control signals accordingly — the prediction function a planner needs to reason about other road users.

The evaluation data is used to evaluate the target planner by computing a reference plan for said time instant based on the scenario state, the scenario state including the ego state at that time instant, and computing at least one evaluation score for comparing the ego plan with the reference plan.— Tools for performance testing and/or training autonomous vehicle planners, US12576864B2

A portfolio centered on verification

The two new grants are not outliers; they sit at the front of a body of issued Five AI patents that concentrate on how an autonomous system is verified and simulated. US12469340B2 covers testing a planning stack by applying driving rules to scenario ground truth and reporting compliance, including rules tied to the operational design domain. US12530284B2 covers a test-visualization tool that renders runs of a driving scenario with a time-varying score against a set of evaluation rules. Together they fence the workflow of running a scenario, scoring the stack against rules, and showing a human the result — the apparatus of validation.

A second strand fences the harder problem of modeling perception error so it can be simulated. US12528479B2 covers fitting noise and misdetection models to training examples to characterize how a perception system mis-sees the world, and US12416921B2 covers a "perception statistical performance model" that produces a probabilistic uncertainty distribution over perception outputs for safety and performance testing. These are claims over how to make a simulator's perceived world realistically imperfect — a prerequisite for trusting simulated test results.

The logic of why a verification company would file those perception-modeling claims is worth spelling out. A simulator that feeds a planner perfect, noise-free perception will pass the planner on scenarios it would fail in the field, where cameras and sensors misdetect and mislocate. The noise-and-misdetection fit of US12528479B2 and the probabilistic uncertainty distribution of US12416921B2 describe how to inject realistic perception error into a test so the planner is graded against an imperfect world rather than an idealized one. Read alongside the rule-checking and visualization grants, they fence a methodology rather than a single product feature: make the test hard and realistic, score the result against rules and reference plans, and show a human where the stack deviated. That is a coherent block of coverage around the question of whether an autonomous system can be trusted, assembled from the testing side rather than the driving side.

Planning under the same constraints it tests against

The portfolio also reaches into planning itself, but framed around optimization and hard constraints. US12547175B2 ("Planning in mobile robots") covers a runtime optimizer that computes a trajectory optimizing a cost function subject to hard constraints, initialized by a trained function approximator. US12536351B2 ("Motion planning") covers searching for an optimal ego action while biasing the simulated agent toward riskier behavior — planning that is explicitly stress-tested against adversarial agents.

Set together, the shape of Five AI's issued coverage is consistent: it concentrates on the verification-and-simulation layer — testing planners against reference plans and rule sets, modeling perception error statistically, visualizing scored runs, and planning under hard constraints — rather than on a shipped perception-to-control driving stack. The two grants that issued this cycle extend that emphasis with a planner-evaluation method and a motion-prediction method. For a reader tracking where an autonomy-software company is fencing its claims, these filings indicate the company is investing in the tooling that proves an autonomous system behaves, a layer distinct from the sensing-and-actuation patents that dominate vehicle-maker portfolios. Each cited patent is now issued and assertable, and collectively they describe enforceable coverage over how an autonomous driving planner is tested, scored and simulated.

The grant dates trace the cadence of that build. The verification and perception-modeling patents issued across 2025 and into 2026 — the planner-testing rule grant in November 2025, the perception-modeling and visualization grants in January 2026, the planning-optimization grants through January and February 2026 — and the two new grants land in March 2026, an issuance pattern consistent with a portfolio that has been steadily converting from application to grant over more than a year. The CPC distribution is consistent too: the testing grants carry B60W 50/06 (conducting tests on subsystems) and G07C 5/08 (recording vehicle operating data); the planning grants carry B60W 60/0013 and B60W 60/0015 (autonomous trajectory generation) alongside G05D 1/02 navigation classes; and the perception-modeling grants reach into G06N machine-learning classes. A reader need not infer intent from a thesis: the classes themselves cluster on testing, recording, modeling and simulating, which is the verification side of autonomy, not the perception-to-actuation pipeline a vehicle maker fences. That is the factual shape of what Five AI has locked in.