Robot space rovers have been exploring Mars for many years, but the hazardous terrain necessitates local path planning to avoid colliding with rocks. This project demonstrates autonomous navigation capabilities in unknown areas (mapless) with a reinforcement learning-based robot behavior learning approach.

As a result, our space rovers can carry out exploration and path planning in dynamic, unknown areas without a map. The learned reinforcement learning-policies are initially trained at large scale with thousands of robot actors in 3D physics simulation end-to-end with 3D sensory input to motor control output. After fine-tuning the policy, it is transferred to a physical rover and tested in the AAU Mars Laboratory at Aalborg University.

Experiments demonstrate that the unified system can achieve a predefined goal while avoiding obstacles in the environment.