Simulations and AI hone the forwarder’s driving performance

How can a forwarder drive as energy-efficiently and carefully as possible through the challenging terrain of the forest? Within Mistra Digital Forest, work is underway to develop trafficability models that can analyse the terrain, identify the best route and plan the movements of forestry machines.

The technology used to develop these trafficability models is based on high-resolution laser scans of the terrain's topography, combined with physics-based simulations. With their help, it becomes possible to train AI models on thousands of scenarios – without them having to drive a single metre in the forest. The project is being carried out by researchers at Umeå University, in close collaboration with Komatsu Forest and SCA.

– Training AI using simulations is an effective alternative when it is too risky and too time-consuming to train models in a physical environment. The strength of AI models is that they can calculate energy consumption, driving speed and the stresses on the machine with great precision, and in a very short time, says Mikael Lundbäck, a researcher at the Department of Physics at Umeå University.

In the route planning of the future – which already takes into account the risk of driving damage and steep terrain gradients – these variables will be able to provide important information for optimal forwarding operations. In addition, they could be a valuable support for inexperienced machine operators.

– Trafficability models are paving the way for the autonomous machines of tomorrow, but right now they are primarily of interest in supporting the industry's efforts to reduce carbon dioxide emissions. How you drive, and where you drive has a huge impact on fuel consumption. This is knowledge that experienced drivers already possess, but which is difficult to formalise, says Mikael Lundbäck.

Calibration work proves more complex than expected 

A large proportion of the work within Mistra Digital Forest involves calibrating the simplified simulation model, so that it better reflects reality. Tests on a physical forestry machine show that a sort of gap arises when AI moves from its virtual training environment into the physical forest, where all manner of disturbances occur. On two occasions, the researchers have collected data from SCA’s logging operations, which the simulation model has then been trained on.

– The calibration work has been more complex than we first anticipated. Reducing the gap is a crucial step in making the AI model’s virtual training environment as true to life as possible. This is a major area of research right now. Progress in this area would streamline the implementation of many different types of AI models, says Arvid Fälldin, PhD student at the Department of Physics at Umeå University.

Dataset and simulation model goes public 

Now, more people can study the impact of terrain on forestry machinery without having to carry out costly and risky field tests. The project has already published open datasets containing high-resolution terrain and machinery data, and a calibrated version of the simulation model will also be made public shortly. The final year of the Mistra Digital Forest project will be devoted to bringing the technology a few steps closer to practical application by training the trafficability model on the calibrated simulation.

- We envisage a forestry industry where data, simulations and AI both streamline the work, and also make it more sustainable, Arvid Fälldin concludes.

Hero: During a field test at SCA in 2024, the forwarder drives over a large obstacle in the form of a 90-centimetre-high rock. Photo: Mikael Lundbäck.