It is already unstoppable. Artificial intelligence has arrived at NASA and it is ready to stay. This new discovery could provoke a technological revolution in the world of space expeditions. NASA has carried out a launch simulation with ChatGPT, thanks to KSPDG (Kerbal Space Program Differential Games). By recreating real conditions and scenarios, the AI was used as a spacecraft pilot in space. Positive results were expected, but the truth is that the AI performed very well, opening doors to its use in real scenarios.
AI in NASA
The Polytechnic University of Madrid and a team from the Massachusetts Institute of Technology (MIT) used a video game simulator to create a perfect scenario in which they tested how an AI would perform as a spaceship pilot. They have used a model such as ChatGPT that has given better than expected results. It has been thanks to the simulator of the popular video game Kerbal Space Program, which is becoming an increasingly used tool for this purpose.
Kerbal Space Program
It is a space simulation video game that allows you to manage a space program. Available for multiple platforms such as Plat Station 5, Xbox One, Wii U, Microsoft Windows, Linux and Mac OS, among others. It is characterized by being extremely realistic, which allows an even more positive user experience. Despite being a video game, it has become a basic tool of the scientific community, as it allows experimentation without facing the costs of real life, as well as the risks. This experiment launched a competition with the aim of testing different artificial intelligence systems. The researchers set objectives such as intercepting satellites, positioning themselves in specific positions and escaping threats.
How do you train an AI to be a space pilot?
To date, we know the methods of training an artificial intelligence; the model is fed with algorithms, physical equations and detailed simulations. It is given as much information as possible so that the answers are as accurate as possible, and it has a bank of information to draw on. However, this was not the case this time. We chose to use only natural language. The model was provided with information about the environment, position, speed and orientation, and after translating it into text, the model also responded as natural language.
Through software, its responses were converted into text that the model could understand. You could say that the model received a text message, and also responded in the same format what to do. The advantage of these models is that they have already been trained with an enormous amount of information, thus saving time in test simulations. In the case of the study, examples from previous missions were provided, and through the technique of “few-shot prompting” (which consists of providing a few examples on which to base the answer), and “fine-tuning” (specialization of the model), much more refined answers were obtained.
Which AI model is in the lead?
The researchers not only made use of ChatGPT, but also of the LLaMA model (created by Meta). Each has features that provide different benefits. Although both were tested under similar conditions, ChatGPT 3.5 proved to be faster, while LLaMA showed a greater ability to learn as the training sessions progressed. The competition held under the name KSPDG, Kerbal Space Program Differential Games, brought together both models and other traditional optimization techniques. The final ranking showed that both were runners-up against a traditional method, revealing a superior capability than expected.
Although the results of the experiment reveal a potential to be exploited, we should not forget that it presents numerous risks. Some of the answers given still made mistakes or nonsense answers, which would be scenarios not only of failure, but also of danger if they were to occur in real scenarios. We will have to wait for the models to refine their answers, but I am convinced that we will see this technology sooner or later.
