Patent grants tend to arrive in clusters, and the cluster is often more informative than any single record. On July 7, 2026, Waymo LLC was issued US12674866B2, a patent directed to detecting and classifying traffic signs by fusing camera and radar data. Taken alone, it reads as one perception improvement. Taken together with the five sibling grants that issued in the same late-June-to-early-July window, it fills in a wider outline: a portfolio that documents the autonomy stack from raw sensor input through downstream driving decisions.

The hero grant is directed to early fusion. Rather than running a camera pipeline and a radar pipeline separately and reconciling their outputs, the disclosed system generates camera features and radar features for the same pixels of a shared coordinate system and the same moments in time, then combines them into a single tensor before a neural network interprets it. The abstract states the approach plainly.

The disclosed systems and techniques facilitate efficient detection and classification of traffic signs in driving environments. The disclosed techniques include, obtaining, using a sensing system of a vehicle a first set of perspective camera images of an environment and a second set of radar images of the environment. The techniques further include generating, using a first neural network, one or more camera features characterizing the first set of images, generating, using a second neural network, one or more radar features characterizing the second set of images, and processing the one or more camera features and the one or more radar features to obtain an identification of one or more traffic signs in the environment.— Detection and classification of traffic signs using camera-radar fusion, US12674866B2

Claim 1 is more specific about the mechanics. A first neural network produces a camera feature per pixel per time, a second network produces a radar feature at the same pixel and time, and the two are fused into a tensor that aggregates across both pixels and multiple times. A third network then processes that fused tensor to output the semantic content of the traffic signs, which the record describes as the sign's type, its value, its relevance to the vehicle, and its location. The classification of the record under CPC groups spanning radar signal processing (G01S 7/417, G01S 13/867) and autonomous driving control (B60W 60/001) reflects that span from sensing to actuation. Radar contributes range and weather robustness that camera-only sign reading lacks; folding it in at the feature level, rather than as a late sanity check, is the direction the grant documents.

The cohort traces a full pipeline

What the surrounding grants suggest is a company patenting the connective tissue of an autonomy system, not just its sensors. Upstream of sign reading sits LiDAR: US12675887B2, also issued July 7, is directed to high-throughput point-cloud processing using a temporal neural network with time-point queries. That record and the hero grant share a common thread — treating sensor streams as tensors organized across space and time, then letting a network extract structure — which points to a consistent architectural preference across the perception layer rather than one-off fixes per sensor.

Move downstream and the cohort turns to prediction and planning. US12668282B2, granted June 30, applies diffusion models to predict the future trajectories of multiple agents at once — the generative-modeling technique family more commonly associated with image synthesis, here repurposed for forecasting how nearby road users will move. US12669341B2, from the same day, is directed to speed reasoning under uncertainty when a vehicle pulls over, including reasoning about occluded parking spots the vehicle cannot fully see. These are decision-layer records: they describe what the vehicle does with a perceived scene, not how it perceives it.

Two further grants sit at the boundary between the two. US12671793B2 is directed to adjusting a vehicle sensor's field-of-view volume in response to detected degradation — a record about keeping the perception input usable as conditions change. US12668277B1 uses collision costs to maintain safe distances, a planning-side constraint that consumes perception outputs. Read as a set, the six grants line up as a sequence: acquire and clean sensor data, fuse and interpret it, predict what other agents will do, and choose speeds and distances accordingly.

What the pattern signals commercially

For a robotaxi operator, the commercial frame is that the moat, if there is one, lives in the integration rather than any single component. The hero grant's insistence on training its three networks together, and on fusing modalities before interpretation rather than after, is consistent with a stack designed as one system. The cohort reinforces that: the same organizing idea — neural networks operating over spatiotemporal representations — recurs from LiDAR processing to trajectory prediction. A portfolio built this way documents an approach to the whole driving task, which is the asset a company selling rides, or licensing a driver, would want on record.

The timing is also worth noting as a signal in itself. Five of the six grants cleared examination inside a single week, which reflects filings made years earlier now maturing into issued patents. That lag means the cohort describes design choices the company committed to some time ago and is only now seeing formalized. The traffic-sign grant's use of radar as a first-class fusion input, the diffusion-based approach to multi-agent prediction, and the explicit handling of occlusion during pull-over maneuvers each read as deliberate bets on how to make an autonomous vehicle robust in the messy middle of real driving — weather, clutter, and the parts of a scene the sensors cannot see. The grants convey where the engineering has been pointed; the market read is left to the reader.