ChatGPT achieves excellent results in spacecraft autonomous control simulation competition

Recently, a study on the use of Large Language Models (LLMs) for autonomous control of spacecraft has attracted attention. Researchers tested the performance of ChatGPT in spacecraft manipulation through simulated competitions. The results showed that ChatGPT achieved excellent second place in the autonomous spacecraft simulation competition, second only to a model based on different equations. This achievement not only demonstrates the potential application of LLM in the aerospace field, but also provides new ideas for the development of future aerospace autonomous systems.

With the continuous development of aerospace technology and the increasing number of satellites, humans will no longer be able to manually control all satellites in the future. In addition, in deep space exploration, the limitation of the speed of light makes it impossible for us to directly control spacecraft in real time. Therefore, developing robot systems that can make autonomous decisions has become an important development direction in the aerospace field. In order to promote innovation in related technologies, aerospace researchers have created the "Kambala Space Program Differential Game Challenge" based on the popular game "Kambala Space Program" in recent years. This challenge provides researchers with a relatively realistic environment for designing, experimenting, and testing autonomous systems. The competition includes various scenarios, such as tasks of chasing and intercepting satellites, as well as tasks of evading detection.

In an upcoming paper to be published in the journal Advances in Space Research, an international research team introduces their competition proposal: a commercial LLM similar to ChatGPT and Llama. The reason why researchers chose to use LLM is that traditional autonomous system development methods require multiple training, feedback, and optimization, while the mission of the Kambala Challenge is to simulate real-world situations as realistically as possible, which means that the task may only last for a few hours, making continuous optimization of the model impractical. In contrast, LLM has already been trained on a large amount of human text, so in the best case scenario, they only need a small amount of careful cue word engineering and a few attempts to find the appropriate context for a specific situation.

According to IT Home, in order to enable LLM to actually manipulate spacecraft, researchers have developed a method that describes the status and objectives of the spacecraft in text form and transmits it to LLM, requesting advice on how to adjust and manipulate the spacecraft. Subsequently, researchers developed a conversion layer to transform LLM's text-based output into functional code capable of operating simulated spacecraft. Through a series of simple prompts and some fine-tuning, researchers successfully enabled ChatGPT to complete many of the testing tasks in the challenge and ultimately achieved second place in the competition.

It is worth noting that these research works were completed before the latest version 4.0 of ChatGPT was released. However, the application of LLM in the aerospace field still faces many challenges, especially the problem of avoiding "illusions" (i.e. meaningless and unreasonable outputs). In real-world scenarios, such erroneous outputs can have catastrophic consequences. However, this research result still fully demonstrates that even existing LLMs can be applied in unexpected ways to practical work after absorbing a large amount of human knowledge.