Tesla Patent Concentrates 3D Perception Detail Where Driving Scenes Need It

An automotive occupancy network faces a basic scaling problem: make its three-dimensional grid twice as fine in every direction, and the system must process eight times as many volume elements, or voxels. A Tesla patent published in March 2026 addresses that growth by changing voxel size according to the detail a region requires, rather than applying maximum resolution throughout the vehicle’s surroundings.

The patent describes 33-centimeter-sided voxels as the default and 10-centimeter voxels for nearby occupied regions. Where a voxel is only partly occupied, the system can subdivide it further to represent features such as curved object contours. The practical objective is selective computation: preserve detail near relevant geometry without paying the processing and power cost of running the entire three-dimensional scene at the finest available resolution.

That distinction matters because a chip’s peak operations-per-second rating does not show how efficiently a perception stack uses its available capacity. Suppliers including Desay SV have characterized 2,000 trillion operations per second, commonly abbreviated as 2,000 TOPS, as an emerging threshold for Level 3 driving systems. That figure is a supplier position, not a legally recognized U.S. approval threshold or an established industry standard.

Resolution is a resource-allocation decision

Occupancy perception estimates whether locations around a vehicle are occupied instead of relying only on a conventional list of detected objects. This can represent road users, barriers and irregular geometry within a common spatial structure. A broad survey of automotive occupancy-perception methods also shows how the field spans camera-only, LiDAR-based and multimodal architectures, with substantial differences in how sensor information is projected and fused.

Uniformly shrinking voxels improves spatial granularity, but the cubic increase in cell count quickly becomes expensive. Going from 33-centimeter cells to 10-centimeter cells across an entire volume would create far more than a simple threefold increase in work because refinement occurs along length, width and height. Tesla’s proposed hierarchy instead resembles an engineering level-of-detail strategy: use coarse representation where little geometry is present and reserve finer cells for occupied areas near the vehicle.

The described architecture extends beyond a single frame. It applies Transformer processing to multiple historical frames and calculates occupancy flow for moving voxels, adding a time dimension to the spatial estimate. Tesla also describes unsupervised training intended to reduce dependence on manually annotated three-dimensional data. Those choices address separate constraints temporal consistency and training-data preparation but they do not by themselves establish production performance or Level 3 suitability.

XPeng prunes the model; Huawei adds sensor diversity

XPeng represents a different integration choice. After removing LiDAR in 2024, the company made occupancy processing central to its pure-vision architecture and paired model pruning with its Turing chip. At a 2025 autonomous-driving workshop associated with the Conference on Computer Vision and Pattern Recognition, XPeng presented its work on scaling driving models.

XPeng has described a mass-produced “2K” pure-vision occupancy network containing more than 2 million grid elements, with an approximately 180-meter perception radius and claimed modeling precision of 5 cubic centimeters per voxel. It has also reported an XNet utilization change on Nvidia Orin hardware from 122% to 9%, about 40% better urban-road recognition accuracy and roughly 20% lower power consumption. These remain company-reported figures; the meaning of utilization exceeding 100% is unclear without a stated normalization method, and the claims should not be treated as directly comparable benchmarks.

Huawei’s ADS 3.0 takes the opposite sensor-system direction. Its General Obstacles Detection Network combines camera, LiDAR and millimeter-wave radar information at the feature level. Sensor diversity can provide greater redundancy, while imposing additional hardware cost, packaging demands and fusion complexity. Huawei-equipped vehicle volume, reported as approaching 1.9 million units by early July 2026, indicates deployment scale but does not resolve those engineering tradeoffs.

For U.S. automotive developers, Tesla’s patent is most useful as an illustration of why accelerator ratings alone are incomplete. Higher peak compute can expand the design envelope, but occupancy resolution, temporal processing, sensor selection and model sparsity determine how much of that capacity is consumed. The unresolved milestone is implementation: the patent defines a way to concentrate spatial detail, but production behavior, power demand and safety performance would still need to be demonstrated in an operating vehicle.

By Robert McKinney — Editor-in-Chief for AMI’s automotive and mobility coverage, with a mechanical engineering background and a decade reporting on powertrain systems, EV innovation, and global vehicle manufacturing.