In-Orbit Control of Floating Space Robots using a Model Dependant Learning Based Methodology

Use of autonomous space robots show promising potential for precise in-orbit proximity operations like in-orbit servicing and debris capture. However, manipulators mounted on board a satellite present a highly complex and nonlinear dynamic system, which is hence difficult to control for precise in-orbit tasks. We had, in our previous work, presented a Non-linear Model Predictive Controller (NMPC) for Free Floating and Rotation Floating space robots in order to design an optimal path that the end-effector can follow while being controlled to reach the target. However, the MPC optimization problem has to be solved online with the requirement of obtaining the solution within the specified loop rate for a stable performance. Due to the high computational time taken by the MPC's optimization routine, the update frequency of MPC becomes a limiting factor when deployed even on moderately complex hardware systems. This led us to modify the existing controller and use a parameterized Neural Network based controller which learns the optimal policy from the MPC solution. Accordingly, in this work, we solve the optimal control problem via Iterative Linear Quadratic Regulator (iLQR) and use it as means to train a Neural Network (NN) policy online. The final control value for the space robot is hence a weighted combination of the control efforts obtained from the iLQR and NN policy. The accuracy of the proposed modification to a conventional Model Predictive controller and its ability to perform the control objective is demonstrated.