Gaby
Research teams competing in training AI models for piloting spacecraft were pleasantly surprised to find that ChatGPT performed remarkably well. To their astonishment, the large language model (LLM) secured second place during competitive testing in this new domain.
“You act as an autonomous agent piloting a fighter spacecraft,” was the first prompt from the researchers, who wanted to explore yet another talent of ChatGPT.
Scientists have long been interested in developing autonomous systems for managing satellites and navigating spacecraft. With the current number of satellites already overwhelming and even more expected in the future, humans won’t be able to manually control each one. This means that we will need to rely on artificial intelligence models to make decisions independently, as reported by Live Science.
What Was the Competition?
To encourage such futuristic innovations, aerospace researchers from various countries launched the Kerbal Space Program Differential Game Challenge. This competition serves as a kind of playground based on the popular video game Kerbal Space Program, designed for the development and testing of autonomous piloting systems in a realistic environment. The AI’s tasks included several scenarios, such as a mission to chase and intercept a satellite, and a mission to evade detection, among others.
In an article for the Journal of Advances in Space Research, researchers presented a potentially successful participant—a commercial LLM like ChatGPT and Llama.
The scientists opted to use large language models because traditional approaches to developing autonomous systems require numerous cycles of training, feedback, and refinement. LLMs are highly effective since they are already trained on vast amounts of human-written text. Thus, they only need well-crafted prompts and a few attempts to grasp the correct context for a given situation.
The researchers developed a method to translate the spacecraft’s current state and objectives into text. They then uploaded this to the LLM, requesting recommendations for the orientation and maneuvering of the spacecraft. Following that, the scientists created a translation layer that converted the LLM’s text prompts into functional code capable of controlling the vehicle (or rather, its simulation).
With a small series of prompts and some fine-tuning, the scientists had ChatGPT complete a series of tests. Ultimately, it achieved second place, while first place went to an AI model based on different algorithms.
Interestingly, the testing took place before the release of the latest, fourth version of ChatGPT. The researchers noted that they have a lot of work ahead in this area, especially when it comes to preventing AI models from experiencing “hallucinations” (unwanted, nonsensical conclusions) that could lead to disasters in real piloting conditions.
Most importantly, the researchers once again demonstrated the limitless potential of AI in absorbing vast amounts of human knowledge and applying it across a wide range of activities.
