Gaby
Teams training AI models to pilot spacecraft were surprised when ChatGPT, a large language model, performed so well it finished second in a competitive test of autonomous piloting.
The researchers began with a simple instruction: “You act as an autonomous agent piloting a fighter spacecraft.” They wanted to see how far ChatGPT’s abilities stretched.
Scientists have long worked on autonomous systems for managing satellites and navigating spacecraft. With the current number of satellites already overwhelming ground teams—and many more expected—humans won’t be able to control each one manually. That means we’ll need AI models that can make independent decisions, Live Science reported.
What Was the Competition?
To push that work forward, aerospace researchers from several countries launched the Kerbal Space Program Differential Game Challenge. The contest is a playground built on the popular video game Kerbal Space Program, designed to develop and test autonomous piloting systems in a realistic setting. Tasks included missions to chase and intercept a satellite, to evade detection, and other scenarios.
In a paper in the Journal of Advances in Space Research, the team described a commercial LLM—like ChatGPT or Llama—as a promising entrant.
The researchers chose large language models because traditional methods for building autonomous systems require many cycles of training, feedback, and refinement. LLMs are useful because they’re already trained on huge amounts of human-written text, so they often only need well-crafted prompts and a few attempts to understand the context for a task.
The team converted the spacecraft’s state and goals into text and fed that to the LLM, asking for recommendations on orientation and maneuvers. Then they built a translation layer to turn the LLM’s textual responses into working code that could control the simulated vehicle.
With a few prompts and some tuning, the scientists put ChatGPT through a battery of tests. It earned second place; first place went to a different AI that used other algorithms.
The tests happened before the release of ChatGPT-4. The team warned there’s still a lot of work to do—especially around preventing AI “hallucinations,” the kinds of incorrect outputs that could cause real-world disasters if they appeared in live piloting systems.
The experiment shows how readily AI can absorb human knowledge and apply it to new, complex tasks.
