Speaker
Description
To properly reconstruct shape and dynamics of a target a chaser has to interact with is a crucial skill that servicing spacecraft shall ensure in performing on-orbit servicing missions. Imaging based sensor suite on board the chaser is the baseline for target data acquisition and drives the proper fly-around guidance synthesis in the mission design phase.
That scheme, however, lacks flexibility and timeliness, highly desirable whenever approaching partially unknown target. A promising solution, here discussed, stays in planning fly-around paths via deep reinforcement learning fed by the set of images acquired during the close proximity operations. To get robustness, fundamental for this application, a vast database of scenarios is adopted for the training phase, together with an artificial neural network settling for the agent policy and the state-value formulation. The core algorithm exploits the recurrent neural networks technique and designs the time history of the maneuvers in order to get the controlled chaser trajectory. The relative guidance is then output learnt to maximize the imaging sequence needed for the mission, taking into account light and background constraints, which might drastically affect the imaging quality and the burden of the data post-processing.
