Researchers develop AI that shrinks robot navigation maps by up to 96%, making exploration algorithms faster and more efficient

A research team created a machine-learning system that automatically removes unnecessary data from graphs (map structures) that robots build while exploring unknown spaces. The system reduced graph size by up to 96% compared to standard methods, while keeping exploration performance stable.

The system uses a transformer (a type of neural network trained with reinforcement learning) to decide which map points to discard in real-time. Instead of storing every location a robot visits, it learns which details matter for navigation and which are redundant—similar to how you'd sketch only the key landmarks on a map rather than every street.

Roboticists and engineers building autonomous exploration systems will benefit: smaller maps mean robots use less memory and compute power, enabling deployment on resource-limited hardware like drones or smaller mobile robots. This matters for search-and-rescue missions, warehouse automation, and space exploration where hardware is constrained.