Robotics versus human decisions and operations

The aim is to equip robot technology with generic capabilities in active perception, world modeling, planning & control and gripping & manipulation. Project leader prof.dr.ir. Eldert van Henten of Wageningen University & Research talks about the current state of affairs.

The program involves three groups of industries, greenhouse cultivation, biscuit and chips manufacturing and poultry processing. For all three use cases, the research program will demonstrate new technologies, which will progressively move forward from university laboratory to industry factory.

Areas of expertise

The Wageningen University & Research (WUR) will lead the way and all universities will bring in their particular areas of expertise. The active perception part is being taken care of by University of Amsterdam and WUR; TU Eindhoven and WUR together tackle the world modeling subject. Planning and control will be the area for TU Delft and TU Eindhoven to excel, while TU Twente and TU Delft will cover the gripping and manipulation chapter.

Variation and complexity

The poultry industry features an enormous amount of technology, but the processing line still requires quite some staff. People particularly work in environments that require multitasking. Precision, product variation and complexity play a role here. Until now dealing with these issues has been a human monopoly. Technology needs to make a considerable step forward to take over these tasks.
“Regarding variation and complexity, robotics in the automotive industry have made gigantic leaps in the past decade,” says Eldert van Henten. “We want to achieve the same progress for the poultry industry. We can use what’s available and further develop that for the agri-food domain.”

Technological capabilities

Several technological capabilities are required to mimic the actions of a human being when faced with product variation and complexity.

1. Active perception. An effective camera system can be used adaptively, just like humans when they turn their heads or turn the object (the piece of meat), to discover what they’re looking for, a strip of fat for example. This is a technological aspect which involves 3D reconstruction as well as understanding of the object and the scene.
2. People learn and store their knowledge, so that they know what to decide next time. When they see a piece of meat, they know where they can expect to find the fat to cut off. People do this all the time; they gather knowledge through active perception and store this existing knowledge in a so called ‘world model’. It is possible for self-learning robots to do their own, similar world modeling and progressing in time obtain a more focused perspective on meat.
3. People just take up their knives and start cutting. Almost intuitively, they know their moves and know when and how to make them. For a robot, however, these actions must be strictly regulated with planning and control. First, the robot has to recognize the situation, then it has to search in its memory for the appropriate actions that go with the situation and then finally execute them. At the moment, this requires a huge number of calculations. Using state-of-the-art learning methods, for instance, the FlexCraft project aims to simplify this task.
4. Gripping and manipulation is about mechanically 'picking up things'. The human hand is a wonder of engineering, very complex for robotics to imitate. TU Delft and TU Enschede have already developed simpler mechanisms which take into account the shape, size, hardness and slipperiness of the object to be picked up. These mechanisms will be easy to maintain and clean.
   
   

Singulation and trimming

Eldert van Henten continues, “Marel set out to robotize three concrete challenging operations in poultry processing: the first one is "gripping and processing of non-singulated products". Singulation is quite a problem in any factory; the equipment used requires a large footprint. This wouldn’t be necessary if a robot could pick single products from a bulk bin.”
The second challenge is the "stylized positioning of poultry meat in a supermarket tray".
The third operation to robotize is "trimming of blood spots or fat from a piece of meat that is held in place". “This is the most exciting one, because humans use both hands to carry out this complex operation. While one hand holds the product, the other one cuts with a knife. But I have high hopes for this one, as our technology partners are particularly eager to find a solution for this challenging problem.”

TRL level

The research project will have a duration of 4 years. By then, it must have reached TRL level 4 or 5. Technology Readiness Level is a NASA standard from 1 to 9, where level 1 represents a brilliant idea and level 9 a system meeting end-product requirements. TRL level 4/5 means that the physical product is being assessed in an industrial environment, close to practice, but not yet integrated in-line.
“After that, the market can take over the robot project or, if needed, additional academic research can be put in place, for example to narrow down, for example, current wide-ranging research to more dedicated practical solutions.”

Co-developing

“It’s important for all parties that the knowledge gained is really applied in industrial practice and doesn’t end up in a drawer. Yet, the program also needs to generate good science. Both aims will go hand in hand.
Right from the start, Marel’s technicians will play a very active role in terms of joined-up thinking, transfer of knowledge and co-developing. Ph.D. students or postdocs will really dive into the food processing practice, visiting processing plants and sitting down with the R&D department. They need to see what happens, find out what the challenges are and where the focus of attention should be. The academics will certainly spend some serious time at Marel,” concludes Eldert van Henten.

Pictures courtesy of NWO FlexCRAFT, WUR