REINFORCEMENT LEARNING FOR SPACECRAFT MANEUVERING NEAR SMALL BODIES

We use neural reinforcement learning to control a spacecraft around a small celestial body whose gravity field is unknown. The small body is assumed to be a triaxial ellipsoid and its density and dimensions are left unknown within large bounds. We experiment with different proprioceptive capabilities of the spacecraft emphasising lightweight neuromorphic systems for optic flow detection. We find that even in such a highly uncertain environment and using limited perception capabilities, our approach is able to deliver a control strategy able to hover above the asteroid surface with small residual drift.