In his research, Daniel Eriksson is particularly interested in the loading task.
“The loading task is simply filling the wheel loader’s bucket with some material from a pile. The material is then dumped at another location – this forms a loading cycle. Filling the bucket may seem easy if you haven’t tried it before, but it is quite challenging to do correctly and efficiently,” Eriksson explains.
Loading is a common task for wheel loaders. It is often repeated during a shift, so automation can help reduce operator workload.
“The bucket filling problem has many automation challenges, the most difficult of which is the ever-changing conditions of a pile and the loading of different materials. This is because the optimal loading strategy depends on the current pile and its properties. Finding a general controller that can handle of these uncertainties is challenging,“ he says.
Artificial intelligence can handle these challenges by using large amounts of training data and learning the best loading technique from sensory inputs on the wheel loader. Eriksson explains that modern AI applications require large amounts of data, and that collecting this data for heavy machinery is more challenging than, for example, scraping images and text from the web, as many other AI models do. According to him, to truly use the benefits of AI, the data must be diverse and realistic, which can only be obtained from real-world worksites and expert operators.
“The AI models are developed by recording data from the expert operators working in the field under normal working conditions. A neural network is then trained to imitate the operator’s behaviour with a certain material. The result is a controller that is able to load a specific material with human-level performance or sometimes even better,” says Eriksson.
This approach works well for a loading a single material but is not as efficient for loading multiple materials. For this task, Eriksson has developed methods to automatically adapt an existing neural network to previously unseen materials. He has also developed a simulator that uses the training data to rapidly test new strategies and techniques.
Eriksson’s work has the potential to help new operators with the loading task and reduce repetitive work for experienced operators. It also has the potential to increase the efficiency of an entire work site, reducing the impact of labour shortage in the construction industry.
Public defence on Friday 1 November
The doctoral dissertation of MSc Daniel Eriksson in the field of automation and machine learning titled Automatic Bucket Filling: A machine learning approach will be publicly examined at the Faculty of Engineering and Natural Sciences at Tampere University at 12:00 on Friday 1 November at Hervanta campus, Tietotalo building, auditorium TB109 (Korkeakoulunkatu 1, Tampere).
The Opponent will be Professor Rauno Heikkilä from University of Oulo. The Custos will be Professor Reza Ghabcheloo from Tampere University.