Medical Embodied Artificial Intelligence in Healthcare Delivery and Clinical Workflows

Medical embodied artificial intelligence (AI) closes the loop between sensing, action and feedback in clinical and care tasks, from procedures and examination to rehabilitation and hospital logistics. Its clinical value therefore depends not only on what a system can do, but on how its action loop fits the task, the people who supervise it and the service responsible for recovery. In this structured narrative Review, we trace the field from image-guided and teleoperated robots to learned policies and foundation models, and survey its technical foundations along the action loop, from perception and task reasoning to planning, control, vision–language–action and world models and safety monitoring, together with data and simulation resources and applications across eight clinical task categories. A 96-source evidence map and targeted technical searches underpin the synthesis, and a separately defined subset of 40 direct empirical studies characterises the settings in which embodied systems have been validated. Evidence in this subset remains predominantly preclinical or feasibility-based and does not establish clinical superiority or long-term safety. We argue that the principal translational bottleneck is not insufficient autonomy alone but a mismatch among task boundaries, machine authority, evidence maturity and health-service readiness. To address this mismatch, we propose a task-bounded framework that links the operational design domain and allocated authority to the evidence required, the conditions for implementation and the mechanisms for monitoring, revalidation and rollback, and we identify technical priorities in grounded foundation models, shared data, uncertainty-aware autonomy and verifiable safety. The framework prioritises clinically relevant comparators, effective clinician takeover and evaluation of the complete workflow, including preparation, supervision and recovery, and offers a basis for appraising deployment claims; its decision value and effects on care now require prospective evaluation.