Paul M. Sutter, a research professor in astrophysics at SUNY Stony Brook University and the Flatiron Institute in New York City, recently participated in a contest where teams of researchers competed to see who could train an AI model to best pilot a spaceship. The results suggest that an era of autonomous space exploration may be closer than we think.
Researchers have long been interested in developing autonomous systems for satellite control and spacecraft navigation. This is due to the increasing number of satellites and the speed of light limitations preventing real-time control during deep-space missions. To encourage innovation, aeronautics researchers created the Kerbal Space Program Differential Game Challenge.
This platform, based on the popular Kerbal Space Program video game, allows researchers to design, experiment, and test autonomous systems in a realistic environment. The challenge includes scenarios such as missions to pursue and intercept a satellite and missions to evade detection. A paper to be published in the Journal of Advances in Space Research describes the work of an international team of researchers who used commercially available large language models (LLMs) to develop their contender.
ChatGPT excels in spacecraft simulation
Traditional approaches to developing autonomous systems typically require many cycles of training, feedback, and refinement, which is impractical for missions lasting just hours. In contrast, LLMs are pre-trained on vast amounts of human text, requiring only careful prompt engineering and a few tries to adapt to given situations.
The researchers devised a method for translating the spacecraft’s state and goals into text, then using the LLM to generate recommendations for maneuvering. This text-based output was converted into functional code to operate the simulated vehicle. With a small series of prompts and fine-tuning, the LLM completed many tests in the challenge, ultimately ranking second place.
First place went to a model based on different equations. Despite the LLM’s success, Sutter emphasized that there is still much work to be done, especially in preventing “hallucinations” — unwanted, nonsensical outputs that could be disastrous in real-world scenarios. Nevertheless, this study showcases the potential of even off-the-shelf LLMs, once they have digested vast amounts of human knowledge, to perform unexpected and highly complex tasks.
This study’s findings mark a significant step toward autonomous space exploration, driven by advancements in artificial intelligence and large language models. However, further advancements and precautions are necessary before such systems can be fully trusted with high-stakes tasks.
Kirstie a technology news reporter at DevX. She reports on emerging technologies and startups waiting to skyrocket.
