Abstract: Robotic on-orbit servicing is an emerging trend. However, space missions often involve unstructured scenarios with targets and environments of irregular physical properties. Traditional methods rely on external aids like markers to structure the scene, which incurs high costs, complex deployment, and poor robustness. This study aimed to develop robotic manipulation technology for the above mentioned unstructured environments. The authors designed an architecture supporting diverse inputs. The embodied perception module integrated vision-language models and zero-shot learning for unstructured perception. The dexterous manipulation module combined dynamic pose servoing with a two-stage compliance control for fine manipulation. This integration enabled intelligent target perception and flexible skill application without complex scene modification. In a typical power drill operation scenario, an arm-hand robot system executed a fully intelligent process including perception, grasping, moving, adjusting, inserting, and screwing. Experimental verification demonstrated millimeter-level precision manipulation in a scene without prior structural modification. The embodied intelligent manipulation method for robots in unstructured environments proposed in this paper can support the development of space intelligent autonomous control technologies for tasks such as on-orbit servicing.

Key words: embodied intelligence, intelligent perception, dexterous operation, robot, unstructured scenario