Uncertainty Aware Model predictive control for free-floating space manipulator based on probabilistic ensembles neural network

Precise control of a free-floating space manipulator (FFSM) is of a great challenge due to the strong dynamic and kinematic coupling between its arms and base. This paper presents a model-based reinforcement learning framework for precise control of FFSMs with dynamics unknown. Dynamic behavior of an FFSM is predicted by a probabilistic ensembles neural network (PENN) trained off-line. The PENN employs probabilistic neural networks to handle aleatoric uncertainty, which is further combined with ensemble method to capture epistemic uncertainty, and used to plan action sequences on-line under the model predictive control framework. Unlike model-free methods which train a particular policy to pursue maximum reward for the corresponding task, this framework allows the same trained PENN to be applied to various tasks with task-specified reward function. Results of numerical experiments demonstrate the fast and robust performance of the proposed framework for both angular and end-effector position control.