A granted patent is enforceable coverage, which makes a cluster of grants issuing in a single week a useful map of where a company has just fenced off territory. In the week of May 5, 2026, five U.S. patents issued to Zoox, Inc., the robotaxi developer Amazon acquired in 2020. Read together, they do not describe a new sensor, a new actuator, or a new vehicle. They describe the machinery a company uses to test and plan how a driverless vehicle behaves — simulation, trajectory generation, and collision avoidance — and that is where Zoox's freshly issued coverage now sits.

The anchor of the cluster is US12617431B1, on risk determination for autonomous-vehicle simulations. Rather than claiming a driving capability, it claims a way of measuring one: replaying recorded driving logs as simulations, isolating the moments where the vehicle came close to another agent, and computing how hard each party would have had to brake to avoid contact.

These metrics may be used to prioritize development tasks, assess performance of the simulation, and assess performance of vehicle control systems.— Risk determination for autonomous vehicle simulations, US12617431B1

That sentence is the tell. The patent's stated use is internal — ranking what to fix and grading the control stack — which marks the grant as coverage over a company's own validation pipeline, the apparatus by which a driverless program decides whether it is safe enough to deploy. Its classification spread (B60W 60/0015 for autonomous-operation safety, alongside B60W 30/0953 and 30/0956 for collision-avoidance control) places it squarely in the testing-and-safety corner of the vehicle-control taxonomy.

What the grants fence off

The simulation theme repeats. US12617392B1 covers simulation buildout with dynamic time intervals — running a simulated trajectory over a period divided into intervals and varying the integration step as the scenario demands, a filing about the mechanics of the simulator itself rather than the car. The two grants together give Zoox issued coverage spanning both how a simulation is constructed and how its proximity events are scored.

The planning side of the cluster is just as coherent. US12617428B2, on an action-reference generation pipeline, claims a model that modifies a baseline trajectory in different ways to output a set of action trajectories usable to define a tree structure, which a vehicle computing device then searches to choose how to act. US12617413B1 covers the contrastive training of object-trajectory encoders and text encoders — jointly training a model on trajectory data and matching text descriptions so it can later classify and predict how dynamic objects move. And US12617400B2 claims a system that learns a human-like desired speed from human driving data and applies it through the vehicle's control system. Across the five, the common subject is the prediction-and-planning layer: forecasting what other agents will do, generating candidate maneuvers, and selecting among them.

Two details inside the cluster are worth pulling out because they show how tightly the grants interlock. The simulation patents and the planning patents are not separate concerns; they are two halves of one loop. The action-reference pipeline of US12617428B2 generates the candidate maneuvers a vehicle could take, and the simulation grants of US12617431B1 and US12617392B1 supply the environment in which those maneuvers are replayed, scored for proximity risk, and graded. The contrastive-encoder grant US12617413B1 feeds the loop from the other side, producing the behavior predictions that tell the planner what the other agents on the road are likely to do. A company that holds issued coverage over both the maneuver-generation step and the simulated evaluation step holds coverage over the full development cycle by which a driverless stack is tuned — a more durable position than a single point claim, because the techniques reinforce one another. The recurrence of B60W 30/09 and 30/0956 collision-avoidance classes across multiple of the five grants confirms that proximity and contact avoidance is the through-line, not an incidental mention.

The footprint and the field around it

What this buys, in business terms, is coverage over methods rather than parts. A patent on a simulation-scoring technique or a trajectory-selection pipeline is enforceable against anyone practicing that method, which is the connective tissue every driverless program relies on regardless of whose lidar or compute it runs. The same-week grant data shows Zoox is not alone in filing here: in that issue week the autonomous-vehicle grant set was led by volume from NVIDIA and Toyota, with Hyundai, Ford and Zoox each among the named assignees, and the most common classification across the whole set was B60W 60/001 — the autonomous-operation control class these Zoox grants sit inside. The field is filing densely in the same taxonomy.

That density is the context for reading the Zoox cluster as a footprint. The company's five grants are concentrated, not scattered: every one touches simulation, prediction, trajectory generation, or collision avoidance, and none claims a sensing or hardware element. For a developer whose product is a purpose-built driverless vehicle with no steering wheel, issued coverage over the validation-and-planning software is coverage over the part of the system that is hardest to demonstrate and easiest to keep proprietary. The grants describe the apparatus that decides, in simulation and on the road, whether and how the vehicle moves.

None of this speaks to deployment scale, to how many cities Zoox operates in, or to whether the coverage will ever be asserted. A grant is a record of what a patent office allowed, not a forecast of commercial outcome, and many issued claims are never litigated. What the week's record shows is narrower and concrete: in a single issue week, Amazon's robotaxi unit added five patents whose claims sit in autonomous-vehicle simulation, behavior prediction, trajectory selection and collision avoidance — mapping its freshly issued footprint onto the software layer behind the vehicle, in a classification space where the rest of the autonomy field is filing at volume alongside it.