Qualcomm Patents a Way to Push LiDAR's Detection Range Further Out
LiDAR sensors on autonomous vehicles have a blind spot: the farther away an object is, the worse the AI gets at identifying it. Qualcomm's new patent describes a way to work around that limit without changing the hardware at all.
How Qualcomm's LiDAR trick sees farther down the road
Today's self-driving cars use laser sensors called LiDAR to build a 3D map of everything around them, but the AI that reads that map is trained to work best when objects are close to the car. Objects at the edge of the sensor's range are harder to classify correctly because the data gets sparser the farther out you go.
Qualcomm's idea is to trick the AI into thinking it's always looking at something nearby. Instead of feeding the whole map to the object-recognition system from the car's point of view, the system picks several virtual viewpoints spread across the map, then re-measures every object's position relative to each of those closer viewpoints. The AI then labels what it sees from each of those fresh angles, and the results are stitched back together into one complete picture.
The result: objects at the edges of the sensor's range get classified using the same conditions the AI was trained on, so accuracy holds up even at long distances. No new sensors required, just a smarter way of processing the data the sensor already collects.
… modifying the points in the point subset for the second origin point so that the coordinate data for the points in the point subset for the second origin point indicate the positions of the points in the point subset for the second origin point relative to the second origin point instead of the first origin point …
Translation: Shifting the reference point for the sensor data to get a clearer picture of distant objects.
How the multi-origin point cloud remapping works
LiDAR sensors emit laser pulses and record where they bounce back, generating a point cloud (a set of thousands of 3D coordinate readings). Standard pipelines feed that entire cloud into a neural network that labels what each cluster of points represents: car, pedestrian, cyclist, and so on. The problem is that the network was trained on dense, close-range data, so its accuracy drops at distance.
Qualcomm's patent describes a pipeline with an extra step. After the sensor collects its point cloud from a single center point (the car itself), the system places several virtual origin points at different locations within that cloud. Think of them as imaginary observer positions scattered around the scene. For each virtual origin, the system pulls out only the nearby points and re-expresses their coordinates as if that virtual origin were the real sensor. That means a distant object that was 80 meters from the car might now look like it's only 20 meters from its assigned virtual origin.
The system then runs the object-recognition model on each re-expressed subset independently, generating labels for each region. A final step combines all those regional label sets into one aggregated label map covering the full sensor range.
The claim also specifies an optimization step: the virtual origin points are repositioned to maximize how much of the sensor's total coverage area they collectively cover, so no part of the scene is left unexamined.
… generating an aggregated label map based on the object labels for the modified points …
Translation: Combining all the processed data to create one unified map of surrounding objects.
What this means for self-driving car safety margins
For anyone riding in or near an autonomous vehicle, longer-range object detection means the car has more time to react. A system that correctly identifies a pedestrian at 100 meters rather than 60 meters can start braking or steering earlier, which is the difference between a near-miss and a collision.
For Qualcomm specifically, Qualcomm's steady investment in automotive perception, this patent fits a broader effort to get its chips and software into the autonomous-driving stack. The technique is hardware-agnostic, meaning it could run on existing LiDAR setups. That makes it relatively easy to deploy as a software update rather than requiring vehicle manufacturers to install new sensors.
That makes this Qualcomm's 32nd filing we've tracked since July in our self-driving sensing work, alongside one on reading road signs and one on pre-mapping roads.
Claim 1 covers any method that collects a 3D map of surrounding objects, plants multiple virtual reference points inside that map, recalculates every object's position relative to those reference points, and runs a recognition program on each resulting section. The claim names no specific sensor, no specific software framework, and no specific vehicle, which means the boundary it draws is very wide.
In practice, that width means any team building software that follows this collect-remap-recognize sequence would be operating inside the territory this claim stakes out, whether they are building self-driving cars, warehouse robots, or delivery drones.
Whether that territory holds depends on whether researchers have already published this same sequence. If they have not, this claim functions as a meaningful toll gate on a practical way to spot faraway objects without spending money on better sensors.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
5 drawing sheets from US 2026/0268685 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →