ChatGPT may be great at giving you recipe ideas with the leftovers languishing in your fridge, but its talents aren’t limited to that. Researchers have wondered whether the artificial intelligence (AI) developed by OpenAI would make a good spaceship pilot. Result: it performs almost as well as a custom-designed system. The future of space exploration will involve robots and AI, it is now inevitable.
“You are an autonomous agent tasked with piloting a pursuit ship”here is the first prompt that a team of scientists submitted to ChatGPT, reports the online media Live Science. Not only did the AI prove itself capable of piloting a ship, but it came in second place in a competition integrated into a piloting simulation.
For years, the idea of entrusting the navigation of satellites and probes to autonomous systems has been gaining ground. Too many satellites, too many missions, too many delays: humans can no longer manage everything, especially when it comes to exploring the confines of the solar system, where the slightest order can take hours to reach the device from Earth.
To stimulate research, engineers created the “Kerbal Space Program Differential Game Challenge”, a sort of sandbox inspired by the famous video game Kerbal Space Programwhere the community can test its algorithms on realistic scenarios: satellite tracking, detection dodging, interception, etc. The objective is to develop an AI that will, ultimately, be able to pilot without human supervision.
ChatGPT, the perfect co-pilot?
In a study to be published, an international team decided to play the ChatGPT card (and its cousin Llama, developed by Meta), AI models already stuffed with human texts, rather than starting from scratch with classic algorithms to be trained laboriously. The advantage? It only takes one prompt well calibrated, a few tries and the AI begins to make spatial navigation decisions on its own.
One of the advantages of this method is its simplicity. Researchers provide the AI with the status of the spacecraft and its mission in text form. She responds by proposing several clear maneuvers: orientation, trajectory, corrections. A module then converts these textual instructions into code to drive the simulated vehicle. Child’s play.
Faced with different situations, ChatGPT passed most of the challenge tests, rising to second place, just behind a model based on specific equations and specifically developed for this type of mission.
Everything is not yet ready today to entrust the controls of a rocket or a satellite to AI. In particular, it will be necessary to resolve the problem of the famous “hallucinations” which, if they are not very dangerous when cooking, could prove catastrophic in space. These results nevertheless open up dizzying perspectives.
With the amount of human knowledge that they have digested, these general AIs could well emerge as the co-pilots of tomorrow, capable of adapting to the unexpected, making decisions in real time and helping us explore the far reaches of the universe.
