GridAI: Drones from a Chinese Startup Fly Under the Forest Canopy Without GNSS or Heavy Compute

网格智算 (GridAI), a Chinese startup founded in 2025, has unveiled an intelligent platform called "GridAI大脑" (GridAI Brain). It lets drones fly autonomously under the forest canopy without a GNSS signal for the first time, measure trunk diameter with millimeter accuracy, and avoid obstacles on their own. The product already ships in small commercial batches.

Why the forest is a "no-go zone" for drones

Under the forest canopy, drones lose navigation. Tree crowns block the GNSS signal, and heavy radio attenuation rules out stable video transmission and real-time manual control. Classic SLAM also fails: it cannot cope with a scene whose texture keeps changing as branches sway slightly in the wind. The industry calls this kind of environment "fully denied" (拒止环境).

Demand, meanwhile, is huge. According to China's National Forestry and Grassland Administration, the country's forest stock reached 209.88 billion cubic meters in 2025, and timber output hit 140 million cubic meters. The UN FAO estimates global annual roundwood harvest at about 4 billion cubic meters, and all of it requires manual stock accounting on cutting plots and manual measurement of trunk diameter.

How GridAI gets by without big data

GridAI builds its understanding of space on "grid-domain learning" (网格域学习) and trains on dozens to hundreds of examples, rather than the tens or hundreds of thousands of labeled images that classic deep learning requires. As a result, compute requirements and hardware power consumption drop sharply.

Instead of recognizing objects from two-dimensional pixels, the system "gridifies" the physical world on three levels:

• Entity grids — discrete, tracked objects in the scene
• Attribute grids — properties that change over time: position, speed, direction of movement
• Relation grids — links between objects (for example, "a nest between two branches")
• Training — 30 photos of nests on branches; on a test of 58 images: 0 false positives, 1 miss, 98.3% accuracy
• The technology's founder is chief scientist 代豪 (Dai Hao), who has more than 30 years of experience

Because GridAI analyzes deep spatial structure and object relationships rather than pixels, its data does not depend on lighting or weather, which makes it more robust to the environment. After the work area is marked out, a drone running GridAI needs no preliminary scanning or model building. It builds a dynamic spatial field in real time, plans its own route, and takes measurements without an operator.

"Only an embodied AI solution that runs offline directly on the device

and delivers high-precision spatial perception and real-time tracking control can handle tasks in the 'fully denied' environment under the forest canopy," said Qian Min, GridAI's marketing lead.

Where the technology goes next

GridAI positions itself not as a supplier of forestry drones but as the creator of a universal "intelligent brain." According to the company, the platform can be embedded in a chip module and applied to robotic arms, driverless vehicles, and underwater robots, giving them the ability to perceive space, make decisions, and act on their own.

The technology has already been tested by the Chinese Academy of Forestry (中国林科院). Its partners include the forestry companies China Forestry Chongqing and Zhejiang Feiliu. In warehouse logistics, GridAI works with Beijing Guangjia and Googol Technology, helping robotic arms with piece picking, packing, and parcel sorting. It has also signed a separate strategic partnership with Shenzhen Unicom.

What this means

GridAI offers an alternative to the mainstream "more data, more compute" approach: spatial perception can be built on small samples and modest hardware, running offline directly on the device. If the approach scales beyond forests, embodied AI will become more accessible in niches that heavy models do not reach because of cost and infrastructure requirements.

FAQ

What is the sub-canopy space, and why is it hard to fly there?

It is the space beneath the tree crowns. The crowns block the GNSS signal, radio communication attenuates heavily, and swaying branches break classic SLAM. As a result, ordinary drones can neither navigate there nor transmit video reliably.

How does grid-domain learning differ from deep learning?

Deep learning needs tens to hundreds of thousands of labeled photos and relies on two-dimensional pixels. GridAI trains on 30–100 examples and analyzes spatial structure and object relationships, so it holds up under changes in light and weather and needs less compute.

Where else can GridAI be applied?

The company names robotic arms, driverless vehicles, underwater robots, and warehouse logistics — piece picking, packing, and parcel sorting. The technology is pitched as a universal "intelligent brain" for different hardware platforms.