AI has almost become indispensable – here’s how it’s (not) used in agrifood
23 October 2024
When speaking to Trim Bresilla, robotics engineer and researcher for Vision+Robotics at Wageningen University & Research, you learn that AI can become everything but a blessing if it is not properly used or even misused. While when properly used, it can really speed up research and development.
Trim Bresilla is part of the ever-growing team of robotics engineers and researchers working at Wageningen University & Research and powering their Vision + Robotics research and development. He specialises in Generative AI or GenAI in short, and deep learning for the purposes of autonomous navigation and operation. Both GenAI and deep learning are meanwhile woven into nearly every aspect of what Vision + Robotics does.
What is GenAI exactly?
Bresilla: “Most people nowadays are familiar with the Chat GPT chatbot from Open AI which is an everyday example of generative artificial intelligence. Based on a few words or a question as input, Chat GPT uses AI to generate a text. While Chat GPT and its alternatives are quite new, GenAI has been used to create images but those models are older. Yet, those models keep getting better and especially since last year, the quality of the models and the outcome have improved. More generally speaking: a GenAI model uses data, numbers or words to generate an outcome. But it is important to realise that the outcome always has a little lower quality than the original input. Because generative models can’t invent data. They have however become very good and reliable.”
The benefits of GenAI in agrifood
“In my domain, field robotics, we deploy GenAI to create imaginable data and situations to make field robots better. Not just in their navigation and operation, but in their overall performance. It’s almost like generating a game engine to test and improve, for instance, navigation in a field of sugar beets whilst detecting each and every sugar beet plant. I did this for example in the Robs4Crops project where we use deep learning for sugar beet plant detection in a generative model. By combining a style transfer of the AgroIntelli Robotti field robot with a GenAI created sugar beet field, we can test and further improve the accuracy and reliability of the sugar beet plant detection algorithm.
And while such models and game engines have come incredibly far and a good model can (almost) reach the same accuracy as a real sugar beet field, you can never exceed the accuracy of the input data. For that, you will always need to combine the AI generated data with real field data. The main benefit of GenAI in this respect, is that you can quickly prototype and test irrespective of the season, the condition of the crop and irrespective of the location. In this particular example, we collaborate with the Agricultural University of Athens and their researchers can trial remotely in Dutch sugar beet fields for instance. That also is a big advantage.”
A screenshot from the Robotic Simulator showing the AgroIntelli Robotti field robot navigating a GenAI created sugar beet field to test the plant detection algorithm. The header photo to this article shows the image generated to make the sugar beet field look realistic.
In another project, Wageningen researchers are using GenAI to create text queries by means of a yield prediction for potatoes. “This is to lead to an online potato yield prediction model based on a farmer’s location, their soil type and their situation. You’d literally ask this ‘potato bot’: What is the predicted or potential yield for field A located along road B with soil type C, and you’d get your answer. We are currently validating the output to assess the reliability of the query. And yet again: the accuracy is determined by the original data with which the model or bot is fed. So we will have to fine-tune the model further with real data to let the base model evolve to a much more advanced and sophisticated model.”
Why link AI with robotics in agrifood?
Bresilla feels that using AI in robotics is important because robots use an increasing amount of vision technologies such as cameras, LiDAR and radar. “The most efficient way to understand the world around you and around a field robot, is to use AI. Both for the detection of plants and other objects, as well as for navigation. Using AI simply offers the best way to understand the world around you. You don’t need any beacons or perimeter wires to geofence a certain area and help a robot navigate. An everyday example is a lawn mowing robot or a feeding robot for dairy cows. These robots used to navigate by staying within a certain perimeter or by following beacons or wires. While nowadays vision technology is used. In combination with AI, the robot can interpret and deal with unexpected and changing circumstances. I like to refer to that as sense and regenerate with the help of AI to refresh the state of a robot.”
A drawback in this respect is that the combination of robotics and AI is not fully exploited in agriculture, because agriculture used to use models that came from other industries such as automotive. “I believe we have to think differently today as with all those startups using the combination of robotics and AI, we have to start setting the trend instead of following it.”
AI to speed up field robot adoption
According to Trim Bresilla, GenAI can also be used to speed up the adoption of field robots. “With GenAI, farmers can simply type a question or use a voice note for field robots to avoid a wet spot in a field. One that wasn’t there yet when he or she planned the operation. ‘Avoid the wet spot between rows X and Y’, could be such a question. The robot’s vision technology will assist in avoiding the exact location of the wet spot. In this way, GenAI can build up trust in autonomous navigation. Another example is to use GenAI to make robots express themselves, to make them speak. Not every farmer is or wants to be a robot scientist. Using GenAI and also deep learning, us scientists can assist in local fields without going there and without being involved directly.”
So why should researchers, companies and startups come to Wageningen for GenAI and deep learning solutions for agrifood? “Wageningen is the best place for this as we not only have computer scientists and specialists in AI and robotics, we also have plant and animal knowledge, and sophisticated growth models of flora and fauna. We use the most reliable synthetic data and have experts in every field on every aspect. That truly is a unique combination you won’t find anywhere else in the world!”
