Looking For Self-Learning Food Robots

FlexCRAFT research program: mimicking the human muscle movements

Work in the food industry can be rather unfriendly sometimes. Temperatures can vary from 3° C in a cold store to a humid 35° C in a greenhouse. In addition, work can be physically demanding and monotonous. Food processors, therefore, have a hard time finding staff. Instead of people, the industry is looking for robots to carry out the tedious work. Such robots would put biscuits in boxes, pick tomatoes or place chicken fillets in a tray.

Shapes, sizes, textures and elasticities
However, discerning the cutting point of a tomato, or picking up a bag of crisps without damaging the contents, poses big difficulties to robots. Many industrial processes already feature robotics, but technical challenges are hindering the agricultural, horticultural and food industries. “Variation is the major bottleneck,” says Eldert van Henten, Professor of Biosystems Engineering at Wageningen University and Research. “If object and action are always the same, there’s no problem for a robot, but unfortunately, no two chicken fillets are equal. The problem with food handling is the wide range of shapes, sizes, textures and elasticities.”

Advancing robots
The large FlexCRAFT research program, uniting five universities and 14 companies, including Marel, has been initiated to advance robotization in the agriculture and food industry. Eldert van Henten has led the FlexCRAFT project since 2019. ”In an earlier stage, people still thought it was technically too ambitious. But by 2018, great progress was made in the automotive and medical sectors in terms of algorithms, vision, sensors, deep learning and control. Based on that, we can now move forward.”

Muscle memory
Eldert Van Henten continues, “Until recently, robots in this industry recorded the environment, deduced relevant data, planned a movement with their robotic arm and started to pick up, cut or manipulate an object. In the end, the robot didn’t know if the action had been successful. That is why FlexCRAFT adds so-called ‘active perception’ to the robot’s algorithm. Now, the robot can permanently acquire new relevant data during the action. Acquired knowledge from previous actions isn’t lost anymore, but stored in a mathematical ‘world model’. The robot becomes self-learning and knows henceforth that it shouldn’t focus on the leaves, but on the truss of tomatoes. From then on, when the same action re-occurs, the robot can retrieve the truss location information from its memory.
In addition, the researchers are drafting a library with stored actions, avoiding a constant complete recalculation for every repeated action. Taking the human hand movement as an example, the TU Delft research team generates mimicked motions for this library. The algorithm looks like a ‘human muscle memory’ that quickly retrieves the optimal way to pick up or grab an object.

Yin-yang
Marel participates actively in the FlexCRAFT project. Allard Martinet, Director Logistics and System Engineering says, “Our plans are to further the robotizing of poultry processing – from cut-up and grading to portioning and packaging. Our RoboBatcher grippers are already able to pick up chicken fillets and drumsticks and position them in a yin-yang pattern in a tray.
However, seasonal products, such as barbecue skewers, with short delivery times, demanded by supermarkets, particularly pose problems for processors. All the more when this involves much manual work, such as marination of cubes, putting them on a skewer and packing them in a tray.
We already apply self-learning algorithms in our machines, but the real difficulty is to cope with the natural differences of products.”

Disruptive thinking
Marel participates in all fields of the research program and even stationed an employee at the WUR university in Wageningen. In addition to optimizations, Allard Martinet hopes to obtain fundamentally new ideas from the program. “When thinking of new solutions, we at Marel start with our own grippers in mind, but independent researchers can come up with extraordinary concepts, such as ​​a harpoon-shaped gripper. It disrupts our own way of thinking and creates new interactions. We expect the program to create various concepts on which we can continue to build, to end up with valuable, practical solutions.”

Connection
Unfortunately, the Covid-19 pandemic has delayed the project. In the last couple of years, many personal meetings have been canceled. Eldert van Henten says, “We always want to achieve as much connection to all researchers as possible. Every six months, therefore, we organize group workshops, hackathon-like, to work on demonstrations. Robotics may look like a matter of programming only, but in fact, real hands-on, on-site work is very much required. Luckily, post-Covid, that has been picked up again. The post-doc students connected to the three use cases have re-involved all researchers from every domain. An exceptional cooperation has now flourished, unseen in other research programs.”