GPS-free localization for defense robots

The context

A global defense company develops and supports ground and aeroterrestrial robots for military and public security missions. These robots—ranging from lightweight reconnaissance units to heavy-duty explosive ordnance disposal systems—operate in environments where GPS signals can be unreliable or unavailable. To maintain operational autonomy, the company sought a solution that would allow its ground robots to self-locate using only onboard cameras.

The logic

The challenge was clear: enable precise localization without GPS by matching what the robot sees to satellite imagery or aerial views captured by drones. While scientific models existed, adapting them required extensive training data. The main constraint was data availability—existing datasets were predominantly urban, whereas the target environments were peri-urban, combining buildings, forests, and open fields.

The solution

After identifying and adapting a suitable computer vision model, the team addressed the data gap by creating synthetic data.

This approach allowed for the creation of custom environments, in this case realistic peri-urban landscapes. We noted that the effectiveness of synthetic data is limited by the creator's expertise. For example, realistic geological features are essential for ground data.

Besides we included military assets such as tanks, absent in public datasets. Each data point could be customized to fit specific requirements.

Each synthetic scene, created in about one day, generated thousands of images at minimal cost.

Next steps in applied excellence

The solution provides a cost-effective, flexible foundation for future enhancements in autonomous navigation.