ChatGPT at the helm: How AI surprised scientists by taking second place in a spacecraft piloting competition

An international team of researchers from MIT, the Polytechnic University of Madrid, and other institutions conducted an unusual experiment: they tested whether ChatGPT by OpenAI could pilot a spacecraft in a simulation. The results, published in a preprint on arXiv (arXiv:2412.11341) and pending publication in the Journal of Advances in Space Research, were striking—ChatGPT placed second in the Kerbal Space Program Differential Game Challenge (KSPDG), where AI agents compete in performing complex space maneuvers. This success suggests that the era of autonomous space exploration may arrive sooner than expected.

How Did ChatGPT Become a Pilot?

KSPDG is an annual competition based on the popular video game Kerbal Space Program, which simulates pseudo-realistic orbital physics. Participants develop autonomous systems to perform tasks such as:

• Intercepting a satellite in orbit
• Avoiding detection
• Maneuvering under time constraints

Traditional approaches to autonomous AI require long training with many iterations, but missions in KSPDG last only a few hours, making such methods impractical. Researchers opted to use a large language model (LLM) like ChatGPT, given its ability to quickly adapt to new tasks. LLMs are already trained on massive text corpora, enabling them to understand context with minimal fine-tuning.

The methodology was as follows:

1. Text-Based Inputs: The spacecraft’s status (coordinates, velocity, orientation) and mission goals were translated into text. For example, the initial prompt was: "You are acting as an autonomous agent piloting a chasing spacecraft."
2. AI Recommendations: ChatGPT proposed how to orient the spacecraft or perform maneuvers.
3. Code Conversion: The AI's text responses were translated into commands for the simulator via a custom software layer.

With just a few prompts and minimal adjustments, ChatGPT completed most tests successfully—falling short only to a model based on differential equations.

Why Does This Matter?

The experiment’s results have far-reaching implications:

• Satellite Automation: There are over 8,000 active satellites in Earth orbit as of 2025, and the number continues to grow. Manual control, as seen with Starlink’s ~6,000 satellites, becomes infeasible. Autonomous AI can handle orbital maneuvers, reducing collision risks.
• Deep Space Missions: For missions to Mars or asteroids, communication delays (up to 20 minutes for Mars) make real-time control impossible. AI capable of independent decision-making is crucial for programs like Artemis or future interstellar probes.
• Versatility of LLMs: ChatGPT demonstrated that language models can tackle tasks they weren’t originally designed for by interpreting text and adapting to new scenarios.

Limitations and Challenges

Despite the achievement, applying LLMs to real space missions faces several constraints:

• Hallucinations: ChatGPT can sometimes produce fictional or incorrect data—dangerous in space. A miscalculated trajectory could lead to a collision orbit.
• Simplified Simulation: Kerbal Space Program uses a two-body model (ship and planet), ignoring complex gravitational influences like the Moon or other satellites. Real missions must account for many variables.
• Data Security: Public LLMs like ChatGPT aren’t suited for critical missions due to potential data leakage. Aerospace companies like SpaceX or Eureka Space use secured corporate AIs.
• Model Limitations: The experiment used a version of ChatGPT prior to GPT-4 (likely 3.5 or earlier). Newer models, such as OpenAI’s o1, could yield better results, but have not yet been tested in this context.

Real-World Applications

ChatGPT's simulation success isn’t an isolated case. In 2024, Italian company Eureka Space tested GPT-4 to control a CubeSat in low Earth orbit. The AI analyzed telemetry, issued commands, and communicated with operators—demonstrating real mission potential. Similarly, NASA has been experimenting with AI for autonomous rover control, such as on Perseverance, highlighting the growing role of machine learning in space.

The Future of AI in Space

The KSPDG results open new possibilities:

• Satellite Fleet Management: AI could optimize orbits for constellations like Starlink or OneWeb, reducing fuel costs and collision risks.
• Interplanetary Missions: Autonomous AI-powered probes could explore Mars, Europa, or the asteroid belt, making decisions without Earth-based input.
• Astronaut Support: LLMs could serve as "onboard assistants," providing guidance or reducing stress during long-duration missions, such as to Mars.

To fully realize this potential, the following are needed:

• Validation Systems: To eliminate hallucinations and ensure reliability.
• Hybrid Models: Integration of LLMs with traditional methods (e.g., differential equations) to boost accuracy.
• Space-Certified AI: Development of specialized AI systems approved for mission-critical tasks.

Conclusion

ChatGPT’s second-place finish in the Kerbal Space Program Differential Game Challenge is more than a curious experiment—it’s proof that language models can handle tasks they weren’t explicitly trained for. By converting spacecraft data into text and translating AI recommendations into executable commands, researchers showed that LLMs can control maneuvers in a simulation. This success, detailed in the arXiv paper, highlights the potential of AI for autonomous satellite navigation and interplanetary missions. While challenges like hallucinations remain, ChatGPT has already shown that the future of space exploration might belong to “thinking” machines capable of surprising even their creators.