A published patent application is not a product; it is a roughly eighteen-month-delayed glimpse of where a company was spending its engineering effort when it filed. That delay is exactly what makes published applications useful as a directional signal — they reveal investment before it reaches a datasheet. For Mobileye Vision Technologies, two applications that published on April 9, 2026 are interesting less for what they cover than for the cluster they sit on top of.

The two fresh filings are squarely in the company's known lane. US20260100056A1 describes a navigation system that uses two cameras to detect traffic lights, with the second camera tuned in a "primary mode" to detect a feature of the light, then acts on the analysis. US20260100051A1 covers a pixel-level approach to vehicle navigation: analyzing individual pixels to decide whether they represent part of a target vehicle, estimating distances to the edges of that vehicle's face, and generating a boundary. Both are camera-perception filings, the discipline the company is identified with, and their classifications (G06V image-understanding classes, B60W 60/0027 autonomous-control) confirm it. The second of the two is the more substantial: rather than drawing a bounding box around a detected car, it works at the level of individual pixels, deciding for each whether it belongs to a target vehicle and estimating its distance to that vehicle's edges before assembling a boundary. That is a refinement of the camera competence, not a departure from it — the kind of incremental perception filing a vision company produces as a matter of course.

The second camera is configured to operate in a primary mode where at least one operational parameter of the second camera is tuned to detect at least one feature of the traffic light.— Systems and methods for detecting traffic lights, US20260100056A1

Where the R&D is quietly heading

The signal sits below those two. Reading the company's recent published-application cluster — a window we widened beyond the single week because the week's two filings are thin on their own — the volume is not in cameras at all. It is in radar and the RF hardware around it. Three applications published March 19, 2026 cover the radar transmit and receive chain directly: US20260079233A1 describes determining a radar transmit-configuration setting based on a road-segment topology; US20260079248A1 covers selecting among transmit configurations with different pulse-repetition intervals; and US20260079234A1 covers a noise-shaping quantizer that reduces bits-per-sample in the digital radar receive path while shaping the quantization-noise spectrum. These are not application-layer claims; they are deep in the signal chain of a radar the company is evidently designing itself.

The antenna and front-end filings reinforce the same direction. US20260088518A1 covers controlling the radiation pattern of a configurable antenna array; US20260081360A1 covers controlling an antenna's polarization based on the angle of an interferer signal relative to boresight. Filing on configurable-beam antennas and interference-aware polarization is the kind of work a company does when it intends to own the radar hardware, not merely consume a third party's sensor. Alongside them, US20260079203A1 describes a system-on-chip functional-safety scheme with parity circuits on the paths between integrated circuits and an on-chip network — a filing that points to the silicon those sensors and models are meant to run on.

A wider sensor footprint, on the record

Two more applications round out the picture and show the perception heritage extending into new modalities. US20260016568A1, published in January 2026, covers determining a predicted behavior from point-cloud velocity information — reading relative movement between elements of a detected target from per-point velocities, the kind of analysis that suits an imaging radar or lidar rather than a camera. US20260038279A1, from February 2026, covers controlling a host vehicle based on a predicted state of a parked vehicle, inferred from a change in the illumination state of the parked car's lights — a camera-perception filing that keeps the vision lineage visible even as the radar work accumulates.

A further application shows the company's interest extending past the sensor and the model into the layer above them — the policy that turns perception into behavior. US20260070584A1, published March 2026, covers a hot-spot detection and reporting system that identifies zones associated with an elevated number of unsafe-condition warnings, collisions or near misses from historical fleet data, then automatically adjusts safety-model parameters such as follow distance and speed settings, or routes the vehicle around the zone. That filing reads as a layer on top of the per-frame perception work: it treats the fleet's accumulated warning data as a map of where the driving policy should tighten. Pairing a map-of-risk filing like this with the radar and point-cloud work below it sketches a stack that runs from the RF front-end up through perception to behavior — the full span a system supplier would file across if it intended to own more than the camera.

Taken together, the published trail points in a consistent direction: a company known for the camera is filing across the full radar stack — waveform, receive-path quantization, configurable antennas, interference-aware polarization — plus the safety silicon to run it. The two camera-navigation applications that published this week are the recognizable surface; the radar and RF cluster beneath them is where the filing volume is. For a business reader, the inference the records support is narrow and grounded: this is a perception supplier investing to widen its sensor footprint beyond vision, and the applications say so roughly a year and a half before any of it ships. What the cluster does not tell us is timing or commercial uptake — a published application is investment disclosed, not revenue booked.