Creative AI research has mostly focused on generating text, images and videos, with limited exploration of live interactive performances. This work examines the challenges and opportunities that emerge when AI systems must operate in real-time, responding dynamically to human collaborators through embodied interaction. Unlike offline generative systems, real-time AI must meet temporal constraints that vary by modality, integrate multi-modal sensing and navigate subjective criteria where novelty may be valued over consistency. The paper introduces Somabotics as an approach that utilises artistic practice to design embodied AI experiences where robots engage human meaning-making through bodily interaction, employing methodologies including performance-led research in the wild and soma design. Emerging research directions include personalisation and long-term adaptation, enabling sustained creative partnerships, cross-modal translation and collective creativity. Advancing this field requires not only technical innovation but also diverse datasets capturing artistic performance, evaluation frameworks appropriate for subjective creative domains and interdisciplinary collaboration between machine learning, human-computer interaction researchers and artists. Creative performance offers unique contexts for exploring how AI systems can meaningfully participate in activities where real-time collaboration, embodied experience and aesthetic quality matter.