Fine-Tuned Language Models as Space Systems Controllers

TL;DR

The paper demonstrates that fine-tuned, relatively small LLMs (ranging from $7$ to $13$ billion parameters) can serve as autonomous space-system controllers across diverse problems, including linear 3D springs, orbital transfers, CR3BP trajectory transfers, and 3-DoF powered descent. By employing Low-Rank Adaptation (LoRA) during fine-tuning, the authors achieve data-efficient training and robust generalization, with the same model able to be fine-tuned for multiple problems with minimal performance degradation. Across linear and nonlinear control tasks, LLM-guided trajectories approach or match traditional optimal-control baselines under many conditions, while offering greater robustness in some nonlinear cases and showing merit in out-of-distribution scenarios. The results indicate that LLMs can act as general space-system controllers, capable of multi-task operation and resilient performance, paving the way for more flexible, language-anchored autonomous space software.

Abstract

Large language models (LLMs), or foundation models (FMs), are pretrained transformers that coherently complete sentences auto-regressively. In this paper, we show that LLMs can control simplified space systems after some additional training, called fine-tuning. We look at relatively small language models, ranging between 7 and 13 billion parameters. We focus on four problems: a three-dimensional spring toy problem, low-thrust orbit transfer, low-thrust cislunar control, and powered descent guidance. The fine-tuned LLMs are capable of controlling systems by generating sufficiently accurate outputs that are multi-dimensional vectors with up to 10 significant digits. We show that for several problems the amount of data required to perform fine-tuning is smaller than what is generally required of traditional deep neural networks (DNNs), and that fine-tuned LLMs are good at generalizing outside of the training dataset. Further, the same LLM can be fine-tuned with data from different problems, with only minor performance degradation with respect to LLMs trained for a single application. This work is intended as a first step towards the development of a general space systems controller.

Figures (15)

• Figure 1: Illustration of the training and control process used in this work.
• Figure 2: Dataset used for the orbital transfer problem: trajectories starting from perturbed initial conditions (black, left), and all trajectories (right), including trajectories starting along the mean trajectory (red), then perturbed.
• Figure 3: Trajectories guided by the Llama-2-7B-1600 (left), and trajectories computed with the optimizer (right), with four non-converging trajectories (red), and 13 trajectories that do not respect the single revolution constraint.
• Figure 4: Position (left) and combined position and velocity (right) accuracy as a function of dataset size.
• Figure 5: One hundred trajectories with biased initial state distribution. Trajectories guided by the Llama-2-7B-1600 (left), and trajectories computed with the optimizer (right), out of which four do not reach the final state (red), and thirteen more trajectories do not respect the single revolution constraint.