Unstung Heroes: Startup’s AI-Powered Tomato Pollinator Gives Bees a Break

by Tony Kontzer

There are nearly a half million acres of greenhouse tomato crops in the world, an area about 35 times the size of Manhattan. In other words, lots of tomatoes.

Growing them requires more than soil, water and sunlight. The plants are self-pollinating, but they need a little help getting the pollen to drop onto the female organ of the flower and trigger the process.

Typically, this is the job of bumblebees, which knock the pollen loose with the vibrations created by their beating flight muscles.

That, however, could change thanks to Israel-based startup Arugga. The company builds AI-powered robots that use computer vision to determine which flowers are ready for pollination and then blast air pulses at them to mimic the action of bumblebees and initiate pollination.

Arugga pollinating robot
Arugga’s new commercial robot will ship later this year.

The performance of the bots is on par with the bumblebees and in some cases better by up to 5 percent, with the potential to perform other tasks and to collect and analyze data along the way.

Competing With Manufactured Bumblebees

Lest it sound like Arugga might throw off nature’s delicate balance, company execs stress that its bots don’t supplant the wild bumblebees buzzing about our gardens.

“The bees we’re replacing are actually manufactured,” said Eytan Heller, co-founder and vice president of business development at Arugga. “They’re grown in labs, packed in hives and shipped with 100 individuals inside. Growers spread the hives around their greenhouses and let the bees do their work.”

In other words, rather than muscling in on a species, Arugga says it is reducing the need for unnatural bumblebee production. Instead of hives of mass-produced bees, Arugga envisions growers purchasing fleets of dependable robots.

Arugga also seeks to work in countries where importing of bumblebees isn’t allowed. In Australia, for example, the pollination process is handled by manual labor. In that case, Heller said Arugga’s robots have demonstrated yield improvements of up to 20 percent, without sacrificing quality.

Riding on tracks between rows of plants, the robots capture images of every flower, and then run Arugga’s AI algorithms to instantly determine each flower’s readiness to be pollinated. Heller said Arugga currently can ensure pollination of one hectare (about 2.5 acres) with 3.5 robots. Next year, it expects an improved algorithm to reduce that to 2.5, and he believes 1 robot per hectare isn’t far off.

Beyond making pollination more efficient, Heller said Arugga’s robots can help growers avoid many issues that arise with bumblebee pollination. For instance, bees can’t work in extreme heat, they can transmit viruses between flowers and they can’t be monitored for precise measurement of their success.

“With our solution, you can monitor and see improvement over time,” said Heller. “The idea is to increase yield within the same space. It’s a big benefit for growers.”

To get that benefit, growers lease Arugga’s robots, paying a monthly fee for each hectare of coverage.

Arugga robot in tomato growhouse
Arugga prototype on trial in U.S. greenhouse.

Arugga only recently launched its tomato pollination robot commercially, working with Australian vegetable grower Costa Group to pollinate the first 10 hectares of its tomato operation. It has also wrapped up a successful trial with a large U.S. grower, with more trials in the pipeline.

NVIDIA Tech for Smarter Applications

Arugga uses the NVIDIA Metropolis platform to develop and deploy its computer vision system. Arugga trains on its own custom dataset, using NVIDIA’s pretrained model YoloV4 with NVIDIA TAO toolkit. It processes raw video from greenhouses with the NVIDIA DeepStream SDK, outsources annotation, and then prepares the training dataset from the annotated imagery.

Each month, the company adds new batches of annotated data to its dataset, and then retrains the model and benchmarks it against previous models. This results in gradual pruning of the model, which is then exported to the robots using TensorRT.

Built upon off-the-shelf carts, Arugga’s robots use NVIDIA Jetson Xavier NX modules, which enable algorithms to run at the edge, allowing the AI magic to occur within the robots themselves.

“I find it very hard to imagine how we could have achieved the same performance without using NVIDIA technologies,” said Ori Shalev, software manager at Arugga. “It’s about being part of the NVIDIA Metropolis partner program and getting early access to NVIDIA SDKs to further enhance and accelerate our AI development efforts. This relationship is super important.”

Next on the Menu: A Fuller Plate

Going forward, Heller envisions adding more AI modules to its robots, empowering them to take on additional tasks. For instance, Heller sees no reason Arugga’s robots can’t detect pests and diseases, treat plants that are showing signs of distress, or apply beneficial solutions to spur plant health.

“We see our solution as a platform to eventually multitask for the grower, a bit like having a Swiss army knife,” said Heller.

He also said Arugga could treat other plants that require similar types of pollination, such as blueberries. Beyond that, there’ll be opportunities he can’t predict.

“We’re excited to be working in this field,” said Heller. “Agriculture in general kind of lags behind other industries in terms of robotics and the ability to do things in a more efficient way.”

That said, the company’s braintrust is being careful not to get ahead of itself, keeping Arugga’s focus squarely on the rollout of its first product.

Said Heller: “There’s plenty on our plate with tomatoes.”

Image credits: Arugga AI Farming Ltd. Company founders Eytan Heller (left) and Iddo Geltner (right) at top.