2024-04-22 05:30:00
From simulation to reality. The United States Air Force organized a real-world dogfight in September 2023 between a two-seater F-16 fighter jet piloted by artificial intelligence (named X-62A) and a human pilot. The robot plane from manufacturer Lockheed Martin did have a human pilot in the cockpit, for safety, but the exercise was entirely controlled by artificial intelligence.
What was the result of the match? A victory for the safety of the device, rejoices Darpa, the agency of the United States Department of Defense dedicated to the development of cutting-edge technologies. Because the pilots of the robot plane would never have had to activate the safety switch during the dogfights. “Integrating a machine that will control a plane in real life is a headache, because the machine then accesses the control commands and can push it to the point of damaging it,” explains Bill Gray, pilot of test at the US Air Force.
VOS INDICES
source
AI’s suicidal behavior upon first encounter
An aerial combat had already taken place between man and machine in August 2020. An AI, developed by the American defense specialist Heron Systems, then achieved the feat of beating a pilot from the American Air Force. The encounter was then entirely simulated, with the human pilot wearing a virtual reality headset.
The experiment revealed the difficulties of placing an AI at the controls of a real plane. The machine’s strategy during the simulation was mainly based on charging the enemy, risking a collision. However, taking the risk of losing a multi-million dollar aircraft is not acceptable for an army. The AI also had a tendency to push the plane to its limits (maximum or minimum speeds), which represents another significant risk of stalling.
A security brick has been added to the AI
The AI developers therefore added a component to the program called Safety trips, making it possible to avoid behaviors that put the device at risk in real life. In total, the X-62A aircraft conducted 21 test flights from Edwards Air Force Base, California, during three separate test periods between December 2022 and September 2023. During these tests, more than 100,000 lines of code would have been modified, before moving on to real combat. Called Vista (Variable-stability In-flight Simulator Test Aircraft), the AI was developed by artificial intelligence specialist Shield AI, which has since acquired 2020 match winner Heron Systems.
An autonomous plane based on machine learning
Apollo lunar mission, autonomous flight of a Black Hawk helicopter in 2021… Autonomous flights are not a new thing. These flight systems are based on type rules: if a given condition is met, then the aircraft has the following behavior. “It’s a very robust system in good conditions,” points out Daniela Rus, research director at the MIT institute.
The task nevertheless turns out to be herculean in the case of an aerial combat: impossible for the developers to anticipate all possible situations. The X-62A aircraft integrates a system based on machine learning. This system allows the machine to take its history into account to make new decisions based on the present and future situation. “It is an extremely powerful system in an environment and situations where flight conditions fluctuate rapidly, making it difficult to clearly and robustly determine rules,” explains Daniela Rus. The X-62A Vista system may have more in common with AlphaGo, the Go game machine, than with the Alias autonomous helicopter system.
Selected for you
1713765396
#Air #Force #achieved #realworld #human #pilot #dogfight
