Deep learning on 3D point clouds for fast pose estimation during satellite rendezvous

Abstract

Light detection and ranging (lidar) is valuable during non-cooperative space rendezvous scenarios. By processing the 3D point clouds, it is possible to provide a navigation solution, consisting of an estimate of the relative pose of the approached spacecraft. To enable a safe rendezvous, the pose estimation has to be precise, but also robust if the output is used as a primary navigation solution. Navigation has to be performed in real-time, and onboard computing hardware has a reduced processing capability. Therefore, the real-time requirement is a main driver of the design. Additionally, a spacecraft often has a symmetrical shape. In this case, the pose estimation method has to account for the fact that multiple attitudes represent the same configuration. This work investigates the use of a point-based neural network, or 3D neural network, for the pose estimation task. This network is integrated in a full pose estimation pipeline, where every component is optimized to achieve real-time requirements on a representative onboard computing hardware. After pre-processing, the neural network produces a relative position and attitude estimation in a single-stage, where the attitude estimation considers the symmetries of the spacecraft. Furthermore, a high-fidelity lidar simulator is used, which enables to generate an extensive synthetic dataset. The method is trained and optimized solely on synthetic data. After training, the pose estimation is evaluated on real lidar data acquired at a hardware-in-the-loop rendezvous facility. Results highlight that the method is accurate and robust, without a loss in performance when evaluated on real data. Finally, the flight-readiness is demonstrated by runtime evaluations on an onboard computer candidate, showing that the method is suited for real-time processing.