Endovascular intervention has become a mainstream treatment of cardiovascular diseases. However, multiple challenges remain such as unwanted radiation exposures, limited two-dimensional image guidance, insufficient force perception and haptic cues. Fast evolving robot-assisted platforms improve the stability and accuracy of instrument manipulation. The master-slave system also removes radiation to the operator. However, the integration of robotic systems into the current surgical workflow is still debatable since repetitive, easy tasks have little value to be executed by the robotic teleoperation. Current systems offer very low autonomy, potential autonomous features could bring more benefits such as reduced cognitive workloads and human error, safer and more consistent instrument manipulation, ability to incorporate various medical imaging and sensing modalities. This research proposes frameworks for automated catheterisation with different machine learning-based algorithms, includes Learning-from-Demonstration, Reinforcement Learning, and Imitation Learning. Those frameworks focused on integrating context for tasks in the process of skill learning, hence achieving better adaptation to different situations and safer tool-tissue interactions. Furthermore, the autonomous feature was applied to next-generation, MR-safe robotic catheterisation platform. The results provide important insights into improving catheter navigation in the form of autonomous task planning, self-optimization with clinical relevant factors, and motivate the design of intelligent, intuitive, and collaborative robots under non-ionizing image modalities.