Abstract: It is a technological development trend in recent years to apply space robot in place of spaceman to perform on-orbit service tasks. Using deep neural network controller, the learning-based space robotic manipulation has shown good potential in adaptability to the unstructured space environments and applicability in fileds such as high earth orbit, extraterrestrial planet exploration, etc. At present, a large number of studies focus on single task robotic manipulation learning problems, for either onground robots or in-space robots. From a new perspective of multi-task learning, a thorough literature review on multi-task robot learning was made, including algorithms and robotic applications therein. To further apply the state-of-the-art multi-task robot learning algorthms, main technical challenges were analyzed and suggestions on key technology development were given. The breakthrough of the above challenges will increase the overall autonomy and robustness level of the space robot system, which is expected to further facilitate the development of China′s on-orbit service towards completely unmanned autonomy.

Key words: space robotic manipulation, multi-task learning, autonomy, on-orbit service, reinforcement learning