A Procedural System to Train Artificial Intelligence to Recognise Satellites
The ASIMOV project aims to design, develop, verify, and validate a space autopilot for mapping non-cooperative objects in low Earth orbit (LEO). Within this joint initiative, our contribution focused on a specific and strategic challenge: the intelligent space autopilot needed realistic training data, in potentially unlimited quantities.
We responded by building a procedural 3D satellite generator, a system capable of automatically composing plausible, geometrically consistent, and industrially credible space configurations.
The Project
Customer need
Every generated configuration had to comply with strict dimensional controls and consistency criteria between components. Not a simple random combination of parts, but a tool that enables the automatic composition of space configurations that are coherent from a geometric, dimensional, and functional standpoint, while maintaining a high degree of controlled variability.
Identified solution
Configurable framework. The solution is based on a framework configurable through external parameter sheets, which allow rules and constraints to be defined without touching the code, ensuring flexibility and scalability. The generation process is guided by a global characteristic size that ensures compliance with the dimensional limits defined by the scenario, while central bodies and components are generated within controlled ranges and with explicit geometric constraints.
The evolution of the algorithm. The algorithm evolved toward a component-first approach, enabling more direct control over the requirements of individual elements and more robust management of the overall composition. In parallel, it was further refined to align with real-world industrial logic, introducing consistency across repeated components, control over models, dimensions and orientation, and support for composite bases.
More components, more variability. Finally, we expanded the 3D library and improved the materials system to increase visual variability while maintaining quality and performance.
A Replicable Model Beyond the Space Sector
The ASIMOV project demonstrates that Game Thinking methodologies are not a privilege of the entertainment world. They are engineering tools applicable wherever there is a need to generate simulated environments, synthetic datasets, or AI training systems. If your organisation faces similar challenges – in defence, robotics, automotive, logistics, or training – the framework we developed for ASI can serve as the starting point for your own solution.
