ChatGPT was badly defeated by Llama. MIT officially announced that AI has a 0% failure rate in flying spacecraft. Musk's Mars colonization is no longer a dream.

[Introduction] The latest research from MIT enables LLMs to directly control spacecraft in a space chase challenge: ChatGPT secures the second place with minimal fine - tuning, while the open - source Llama outperforms it. By using prompt engineering, it can precisely track satellites and save fuel, achieving a 0% failure rate. This validates the efficiency of small - data AI and the feasibility of autonomous spaceflight, paving the way for future space exploration.

Just now, a new research on AI "piloting" spacecraft has gone viral as soon as it was released!

In a space challenge derived from the Kerbal Space Program, research teams including MIT used ChatGPT as the "main control" agent and surprisingly won the second place.

This competition is regarded as an important testing ground for the autonomy of space exploration, and the performance of AI indicates that "AI piloting spacecraft" may be closer than we think!

Perhaps, the space odyssey in 2027 can really become a reality! Satellites and space debris orbiting the Earth may all be autonomously operated and managed by AI in the future.

The research team didn't conduct extensive model training. Instead, they skillfully used prompt engineering and minimal fine - tuning to enable ChatGPT to successfully handle complex space tasks such as chasing satellites and evading detection.

The entire system consists of three steps: text state input → language model decision - making → code execution, demonstrating the powerful generalization and adaptability of the LLM model.

The research paper by MIT and the Technical University of Madrid has been accepted by the Journal of Advances in Space Research and is about to be published.

Paper link: https://arxiv.org/pdf/2505.19896

Quick overview of research highlights:

• ChatGPT uses text instructions to complete spacecraft navigation and control decisions, performing far better than expected;
• The research doesn't require large - scale training and makes full use of the existing knowledge and language understanding of LLMs;
• Although there are still risks such as "hallucinations", autonomous spaceflight has changed from a fantasy to a feasible path.

AI autonomously pilots spacecraft into space

Researchers have long been committed to developing autonomous systems for satellite control and spacecraft navigation.

In the future, there will be too many satellites for humans to control them all manually.

For deep - space exploration, the limitation of the speed of light means that we can't directly control spacecraft in real - time.

If we really want to expand in the space field, we must let robots make decisions on their own.

In recent years, aerospace researchers have created the Kerbal Space Program game challenge to encourage innovation.

This is a test - bed based on the popular Kerbal Space Program video game, allowing the research community to design, experiment, and test autonomous systems in a (to some extent) realistic environment.

The Kerbal Space Program (KSP) was originally a space - flight simulation video game developed by the Mexican studio Squad and released in 2015.

Although it is a game, by adding mods, it can be used as a simulation environment, and these mods can add new features such as more realistic physics.

Although KSP doesn't provide a perfect simulation of reality, its accurate orbital mechanics mechanism has been praised, and it has even established a partnership with NASA, elevating its status beyond that of an ordinary game.

The simulation environment is limited to a two - body problem and only a small number of planets, most commonly a single Earth - like planet named Kerbin.

The challenge includes multiple scenarios, such as tasks of tracking and intercepting satellites and tasks of evading detection.

Researchers decided to use LLMs because traditional control methods require multiple rounds of training, feedback, and improvement.

However, the essence of the Kerbal challenge is to be as realistic as possible, which means the tasks only last for a few hours.

Therefore, continuously improving the model would be impractical.

LLMs are so powerful because they have been trained on a large amount of human - written text.

So, in the best - case scenario, they only need a small amount of well - designed prompt engineering and a few attempts to obtain the correct context for a specific situation.

But can such a "conversational" model really pilot a spacecraft?

Piloting spacecraft with GPT and Llama

First, let's introduce the problems to be solved in the KSP challenge. The agent controls the movement of the spacecraft on all three rotational axes (yaw, pitch, and roll) through thrust engines.

The actions are expressed in the spacecraft's reference frame, including the thrust magnitude of each axis and the duration of thrust application.

The KSP challenge includes the following three scenarios:

• Pursuer - Evader: The agent controls the pursuer. The main goal is to minimize the distance between the pursuer and the evader.
• Target Guarding: The agent controls a hijacker spacecraft to approach another spacecraft.
• Sun Occlusion: The agent aims to position the spacecraft between the evader and the sun.

This research only focuses on the Pursuer - Evader scenario.

In different scenarios of the pursuer - evader game, the initial orbit of the evader remains the same in all scenarios, while the initial orbit of the pursuer varies.

The pursuer and the evader have the same spacecraft parameters.

The evaluation metrics include the distance (in meters) between the pursuer and the evader, the speed (in meters per second) at the closest approach, the fuel consumption (in kilograms) of the pursuer, and the elapsed time (in seconds).

Now that the task is clear, let's see how GPT and Llama pilot the spacecraft. This research explores two approaches:

1. ChatGPT + Fine - tuning

GPT was chosen because it is easy to use, and the focus is on the fine - tuning model strategy.

2. Llama + Prompt engineering

Llama was chosen because of its community support and open - source flexibility. Prompt engineering is used as the main research approach, along with simple fine - tuning.

The researchers developed a method to translate the given state of the spacecraft and its goals into text form.

Then, they passed it to the LLM and asked the LLM to provide suggestions on how to adjust and control the spacecraft.

The researchers then developed a translation layer to convert the text - based output of the LLM into functional code that can operate the simulated spacecraft.

The research generated examples of multiple orbits to collect training data for Llama.

The research also provided the pseudo - code for orbit data generation.

Through a series of short prompts and some fine - tuning, the researchers enabled ChatGPT to complete many test tasks in the challenge, and it eventually won the second place in a recent competition.

Llama beats GPT

Interestingly, OpenAI's fine - tuning API requires customization, but the tools it provides are very limited, especially compared to Llama.

Therefore, the training effect of GPT depends largely on the quantity and quality of data, as well as certain adjustments (especially hyperparameters), among which LRM has the most significant impact.

The application of the chain - of - thought method significantly improves the generalization ability of spacecraft piloting technology in the pursuit problem and successfully guides the model to achieve a 0% failure rate during execution.

The results of Llama far exceed the researchers' expectations.

This model not only follows a stable prograde orbit but also outperforms almost all other methods in the KSPDG challenge.

It should be emphasized that the basic Llama model achieves better results than the GPT model.

However, considering that Llama - 3 is a model competing with GPT - 4, rather than the G