He would still just walk into the flames.

On its four stilts, "Byte" moves swiftly. "But if we let the autonomous robot run toward a fire now, it would simply run through it and be destroyed," says Sören Pirk. The computer scientist leads the "Wildfire Twins" research project at Kiel University. With two million euros in EU funding, he aims to teach machines to move autonomously through forest fires and extinguish fires in a few years with the help of artificial intelligence.

However, a deployment in one of the current forest fires in Southern Europe comes far too soon for the robot – fire departments in Spain and Greece, for example, have been heavily in demand recently. The computer scientists are currently conducting basic research. At first glance, the simulation images on Pirk's screen are more reminiscent of video games. Using the software, the researchers allow trees, undergrowth, or entire forests to burn.

"We need to generate data in our simulation that looks like it comes from a real forest fire scenario," says Pirk. "The whole thing should look photorealistic, like a real forest fire. It's similar to a video game." Only more realistic.

Video cameras alone don't help the team. Images provide the machine with too little information about the fire. "Currently, the robot doesn't know what to do when it detects flames," says Pirk. "It simply doesn't have a template for a solution. It doesn't know whether to extinguish the flame in front of it directly or what distance it should maintain."

The researchers want to use a virtual training environment to teach the AI to find safe paths through realistic fire scenarios. "While I can easily construct a forest from a satellite image, things are less straightforward with the undergrowth," says Pirk. Satellite images don't provide any information about whether the fire can be extinguished, for example. However, this is relevant for a plausible forest model. "That's why we're working on mathematical models and, similar to a computer game, building 3D models of individual trees, the undergrowth, and grasses."

Currently, four young scientists are working on the project alongside Pirk. The training robot "Byte," which costs approximately €100,000 and weighs 25 kilograms, will eventually provide firefighters with detailed information about fire scenarios. The training simulation requires a large number of photorealistic images of forest fire situations, which the AI will then use to learn.

"We hope to have a virtual training environment in five years. However, Byte won't be able to carry out firefighting missions by then," says Pirk. The East Frisian native came up with the idea for his research project in the USA, where he worked for a tech company.

Practical experiments are also necessary. At Schleswig-Holstein's State Fire Service Academy in Harrislee near Flensburg, "Byte" collects data on fires of varying intensities. Once fully developed, the technology could conceivably be used for vegetation fires, for example, says group leader René Heyse: "We're seeing more and more of them."

Heise demonstrates in a fire container how smoke can ignite during a fire. "We're first trying to understand the fire with the robot," says Pirk. The AI has to learn to interpret the flames. "Just like we humans learn not to get too close to a fire."

Fire expert Heyse hopes the technology will help with fire detection. "Where is the fire coming from, in which direction is it spreading, and are there people present?" AI could potentially recommend initial measures based on its ability to calculate how the fire is developing. An AI-supported platform—drones or robots—could also be a valuable aid in preventing emergency personnel from being put at risk.

"The crucial phase is really the situation assessment phase," says Heise. This has to happen quickly in an emergency. AI can help identify areas at risk of collapse or extreme fire phenomena such as layers of hot smoke that threaten to ignite. "If an AI-supported platform can tell me all of this, I'll obviously have much more time outside, for example, to identify people at the window and initiate immediate action accordingly."

In the long term, computer scientist Pirk can also imagine autonomous systems that can independently combat fires - such as those currently raging in southern Europe.