Here AI is driving a forestry machine - for the first time

2023 will go down in history as the year when AI was on everyone's radar. While most people were familiarising themselves with chatbots, a Mistra Digital Forest research project was trying to have AI drive a forestry machine for the first time. Although the transition from simulation to reality was not without its problems, it was a successful test run.  

Anyone who browsed scientific journals such as Nature and Science last year could read about AI flying drones, and navigating robot dogs in difficult terrain. The phenomenon was also on the rise in forestry, and within Mistra Digital Forest tests in which AI drove a forestry machine were carried out for the first time. 

The research group in Digital Physics at Umeå University is behind the project. They have used a form of AI called deep reinforcement learning to train a self-driving pendulum arm forwarder. The development work involves the researchers creating a virtual copy of the forest machine and the forest terrain. By letting the simulated forestry machine drive around in the realistic virtual environment and learn from its mistakes, it gets better and better. In 2023, the AI was sufficiently well trained that it was time to test whether it could drive a physical forestry machine. Viktor Wiberg, then a PhD student at the Department of Physics at Umeå University, supervised the tests, which took place at Skogforsk's Troëdsson Forestry Teleoperation Lab. 

- The tests show that it is possible to train an AI model, first in a simulated environment and then transitioning to letting it drive a physical forestry machine. This is not a particularly easy task because the vehicle has a complex hydraulic system, and in addition, it must be able to navigate rough terrain, says Viktor Wiberg. 

Challenge going from simulation to reality

The focus of the research study is the transition from simulation to reality. 

- This is a relatively unknown area and we need to understand what is crucial in order to successfully manage the transition. No matter how good our simulations are, reality presents challenges that we then have to deal with, says Viktor Wiberg.  

Three tests were conducted a few months apart. It was immediately clear that the physical forwarder interpreted the control signal in a different way to its virtual counterpart, resulting in a jerky and unpredictable pattern of movement. To address this not entirely desirable behaviour, the researchers had to retrain the AI by mimicking the physical forwarder's steering signal in the simulation, thereby more closely matching simulation and reality. Another obstacle in the transmission was the delay in communication, similar to the lag in a car's GPS. This delay does not occur in the simulation and in order for the forestry machine to move smoothly, the AI had to learn to take that disturbance into account.

- What was special about these tests was that we had to adapt the pendulum arm function so that it could be controlled manually. Normally, when we test different remote control solutions, they are managed by the machine's own computer, but now the AI was going to take over the control, says Tobias Semberg, civil engineer in operational systems at Skogforsk and active at Skogforsk's test bed.  

After re-training the AI on the supercomputer in Umeå, the pendulum arm forwarder was able to navigate the test environment on the third attempt. 

- It is impressive that it actually worked, and that we could see so clearly how the AI performed better from session to session. It shows how important it is to start testing on real machines and then go back and redo, in my experience the problems that emerge are rarely the ones you expected initially, says Tobias Semberg.  

"Now we know that it is both cost and time-efficient"

The research team at Umeå University has further advanced this work by investigating how deep reinforcement learning can be used to automate crane operations.  

- We now know that it is both cost and time-efficient to train AI in a simulation initially before moving on to a physical forestry machine. Deep reinforcement learning has great potential in this area, concludes Viktor Wiberg.