Researchers tested how effectively large language models could control a spacecraft, and the result surprised even skeptics: ChatGPT took second place in an international autonomous piloting simulation competition.
Researchers tested how effectively large language models could control a spacecraft, and the result surprised even skeptics: ChatGPT took second place in an international autonomous piloting simulation competition.
As reported by Live Science, the tests were held as part of the Kerbal Space Program Differential Game Challenge. The task was to simulate complex scenarios — such as intercepting a satellite or evading detection. Both classical algorithms and LLM models such as ChatGPT and Llama were used to participate.
The researchers chose the language model because of its ability to quickly adapt to new conditions using text prompts. Instead of the many learning stages typical of conventional autonomous systems, ChatGPT received a description of the current situation and mission objectives in text form. The model then generated instructions for maneuvering the ship, which a special module converted into machine code that controlled the simulator.
After several rounds of prompts and adjustments, the model was able to successfully complete most of the scenarios. It was only inferior to a model based on specially developed physical equations. It is noteworthy that the experiments were conducted before the release of the latest version of GPT-4.
The authors of the scientific paper, which is being prepared for publication in the Journal of Advances in Space Research, emphasize: LLM models can become a useful tool in future space missions, especially where signal delay makes operational control from Earth impossible.
Autonomous control systems are key to scaling satellite missions and deep space exploration. Traditionally, complex algorithms have been developed for this, requiring significant time for training and tuning. The use of generative AI opens up new approaches that are much more adaptive and cheaper. The problem of model «hallucinations» still remains, but the potential is clear: even off-the-shelf LLMs are capable of controlling complex systems given the right context.
As a reminder, we also published an article about how some ChatGPT subscribers have a new feature called «Study Together.» It changes the chatbot’s usual behavior: instead of ready-made answers, it asks clarifying questions to help the user learn on their own.
