Duration: 36 months
Planning and control problems are becoming ever more challenging due to the increased complexity of modern space missions (e.g., re-usable launchers and celestial hopping systems). Classical model-based control methods are reaching a point where they will no longer be able to fulfil performance requirements for such complex systems, thus compromising mission safety and success. To maintain system-level simplicity without sacrificing performance, guidance, navigation and control (GNC) systems must be more versatile and adaptable to the many tasks and system configurations. Adaptability, and thus improved safety, can be achieved by online optimization. Model-free/learning-based approaches [1] remove the necessity of complex, task-specific, high-fidelity models, thus making them more versatile to different use-cases. This project focuses on applying a model-free, online optimisation control algorithm (e.g., learning [2] [3]) to spacecraft landing scenarios. In order to guarantee robustness in the resulting controllers, the algorithm will utilise perception-based control to learn the system states from live image data then subsequently optimise the controller gains with certain robustness and performance bounds [4]. To test the model-free, vision-based GNC approaches online in a realistic-setup, they will be implemented on a multi-copter test-bed. The multi-copter can accurately simulate various space systems in different environments. Modularly removable flexible pendulums, tanks and masses, incorporate varying dynamics (sloshing, flexibility, instability, etc.) [5] [6]. Direct thrust control simulates different gravitational fields as well as thrust vectoring control of the vehicle. And, modular sensors provide the visual input to the GNC algorithms. Overall, the work will provide valuable insight into using model-free approaches on autonomous space systems, simulated on a demonstrator more versatile than those developed for NASA [7].
