(Semi-)autonomous robotic assistance in minimally invasive surgery has the potential to alleviate surgical staff shortage and decrease the workload of medical professionals. These robots must execute complex tasks within unpredictable and unstructured environments encountered during surgery. Although imitation learning approaches have the potential to learn complex surgical skills, the interpretation of robot behavior during safety-critical scenarios, such as surgery, remains a challenge. Through combining interpretable 3D point cloud feature vectors based on domain knowledge with feedforward neural networks and probabilistic movement primitives, domain knowledge-informed movement primitives effectively learn surgical skills while improving the interpretation of robot behavior. The evaluation on test data proves that the proposed method can effectively learn surgical skills based on a small number of demonstrations. Using the proposed imitation learning method, a semi-autonomous robotic assistance for directed gallbladder retraction is introduced and evaluated during gallbladder removal on a silicone liver phantom and ex vivo porcine livers. Achieving over 91 % and 92 % successful gallbladder retractions, the robotic assistance enables effective support for surgeons during these surgical interventions.