Developing unique robotic systems for automated fresh salad ingredient packaging
14 April 2026
In Wageningen, Vision+Robotics experts are developing vision-based robotic picking and packaging solutions that could prepare your next salad box without human intervention. Aneesh Chauhan shares the team’s latest updates and challenges.
It is no secret that the researchers and scientific teams of Wageningen University and Research (WUR) and Vision+Robotics (V+R) experts have gained extensive experience in robotic picking and packaging over the past decades. While previous projects mainly focused on grippers and flexible robotic arms for various salad box ingredients, the current project KB robotics – Use case Food factories focusses on 3D vision-based dosing of ingredients. Dosing involves picking ingredients to meet a requested target weight specification, which varies across different salad recipes. This is a challenging task, as salad box ingredients vary a lot in size, composition and in density and hardness. Aneesh Chauhan, Senior Scientist Computer Vision and Robotics at WUR and V+R, elaborates on the specific challenges involved.
Huge variation in ingredients and composition
“In this project, we are trying to develop and build robotic solutions that can compose a wide variety of salad boxes. We all buy these ready-to-eat salad boxes at supermarkets and in convenience stores. The next time you buy and eat one, have a look at the ingredients in your favourite box. You’ll find various ingredients such as lettuce, beans, cucumber pieces, pepper cubes, potato wedges, rice and pastas like fusilli. All of these packages are made to specifications based on pre-defined salad recipes, including weight of each ingredient. Now picture one single type of robotic arm, picking the exact amount of ingredients required, without physically weighing each portion and still adding the correct weight to the box. That’s what we’re trying to achieve, and I can assure you that this is challenging! Why? Because first of all, we have to deal with many different types of products, and even for a given product, the robot should ‘know’ how to pick the desired weight. We gained experience with picking cut bell pepper cubes in the industry supported NXTGEN – Ready Made Meals project and are currently using that experience to pick kidney beans, fusilli and mozzarella or cheese balls. Since September last year, we have been researching whether we can apply the learnings and methods from that project to our new KB robotics factory use case.”
Humans can quantify weights naturally
Common food industrial practice when composing ready-to-eat salad boxes involves a wide variety of ingredients and materials. Multiple funnels feed and balance the required weight of lettuces for instance. “This is already highly automated, but many products still have to be added manually. Multiple people scoop the required amount and can decide intuitively whether the scooped quantity matches the weight dictated by the recipe. But what happens when the recipe changes and 25 instead of 40 grams of product has to be added? Is intuition still sufficient? We are trying to find out which parts of this process can be automated to reduce reliance on manual, repetitive tasks. The main challenge is how to enable a robot to add the correct amount of product time and again, without weighing the quantity every time the robot scoops or picks a product. Counting individual particles also is out of the question, as the product can be very dense (think of rice, beans or pasta) inside the boxes and crates. The surface from which the pick is made changes each time you grab a certain amount of product. At the start of a new batch, you probably dig into a pile, while at the end, you try to pick products from alongside a hole. In some cases, such as with bell pepper cubes and mozzarella or cheese balls, the product can also be partly submerged in liquid. The key lies in machine learning.”
Measuring weight with 3D vision
Despite their vast experience, it was still quite challenging for the researchers to develop a solution that can measure weight simply by looking at a dense collection of food products. “That is fundamentally a very difficult problem, even if we are at the forefront of this kind of research and development”, Chauhan emphasises. “We rely on a combination of modern vision technologies, machine learning and artificial intelligence (AI) algorithms to achieve our goals. The set-up includes a Zivid-2, a 3D camera and the flexible, lightweight and fully adaptable Piab piSOFTGRIP four-finger gripper. The gripper is mounted on a Universal Robots UR5 6-axis robotic arm, while the camera is fixed at the top of the frame to provide a clear view of the product tray. For training purposes, we use a weighing scale to correlate a visually determined amount of product with the required weight indicated in the recipe. We use machine learning to determine how deep a gripper should ‘dive’ into the box or container to pick the correct amount and weight. To do this, we extract informative features from the 3D point cloud of the pick location and train a series of machine learning algorithms to model how deep the gripper should go into the product to pick the desired weight. One of the main aspects of this is how to take a point cloud – the captured 3D points on the surface of the surface layer of the product – and extract meaningful data we can train a model on. The top of the surface, for instance, has a lot of influence on how deep the robot should pick to reach the desired weight. Machine learning plays a key role in generalising across changes in the surface layer. So far, we have already achieved promising results for kidney beans and fusilli. We currently are at TRL 3 to 4, and our goal is to expand the portfolio of products and to find out for which types of products our solutions and approach work. We have learned that with products that cannot be packed densely and contain a lot of air between the particles, such as mozzarella or cheese balls, it is difficult to obtain a consistent weight using our approach.”
Validate and scale next
This use-case within the KB Robotics project started in June 2025 and will run for at least two more years. It is primarily funded by the Dutch Ministry of Agriculture, Fisheries, Food Security and Nature (LVVN) i via the Knowledge Base program Autonomous Robotics, grant number KB-53-000-010. “This is a research project, and our aim is to see wether our developments can scale to new products and eventually fulfil industrial needs. We do this by validating our models and checking their consistency and reliability in new scenarios and for other types of robotic applications as well. We also want to demonstrate the relevance of our research to the industry. In that respect, Vision+Robotics remains the go-to partner when it comes to expertise in relevant disciplines such as computer vision, mechatronics and agrifood. This combination of technical expertise and domain knowledge is highly complementary and truly unique to WUR”, Chauhan concludes.
The surface, where the pick is made, changes each time you grab a certain amount of product. In some cases, such as with bell pepper cubes and mozzarella/cheese balls, the product can also partly be submerged in liquid. The key lies in machine learning.
From that 3D point cloud of the pick location, we extract informative features to train a series of machine learning algorithms to model how deep the gripper should go into the product in order to pick the desired weight. The results is a depth map like this.
We extract informative features from the 3D point cloud of the pick location and train a series of machine learning algorithms to model how deep the gripper should go into the product in order to pick the desired weight.
