Medical Embodied AI survey: Perception, decision, action

Cheng Zhang and eight coauthors submitted "Towards Next-Generation Healthcare: A Survey of Medical Embodied AI for Perception, Decision-Making, and Action" to arXiv on 5 Apr 2026. The 19-page paper (9 figures) argues that while foundation models have boosted many medical tasks, their limited ability to perceive, understand, and interact with the physical world constrains performance in safety-critical clinical workflows.

What does the survey cover?

The paper systematically surveys the core components of medical embodied AI, emphasizing the coordinated integration of perception, decision-making, and action. It reviews representative medical applications and relevant datasets, analyzes major challenges encountered in real-world clinical practice, and outlines key directions for future research. The authors present this material across 19 pages and 9 figures and link an associated project via a provided URL.

How do the authors frame "medical embodied AI"?

Medical embodied AI is framed as a physical-interactive paradigm that lets agents operate in complex medical environments, bridging perception, reasoning, and physical execution. The abstract contrasts this paradigm with foundation models, noting that foundation models deliver "impressive performance" on many medical tasks but lack the ability to perceive and interact with the physical world, which the embodied approach addresses.

Which applications, datasets, and components are highlighted?

The survey explicitly reviews representative medical applications and relevant datasets, tying them to three coordinated system-level components: perception, decision-making, and action. Perception covers sensory input and scene understanding; decision-making covers safety-critical clinical reasoning; action covers physical execution in clinical contexts. The paper positions these three elements as integrated parts of end-to-end systems rather than isolated modules.

Why it matters

The authors argue that clinical environments couple safety-critical decision-making and physical execution, a coupling that foundation models alone cannot resolve because of limited physical-world interaction capabilities. Framing research around embodied agents that integrate perception, decision-making, and action addresses gaps that prevent current models from operating safely and effectively in real-world clinical workflows.

What are the paper's concrete outputs and signals?

The submission metadata provides concrete signals: arXiv identifier arXiv:2606.15647, submission date 5 Apr 2026, file size 10,121 KB, and the author list led by Cheng Zhang with coauthors Qing Cai, Xingzheng Wu, Xun Yang, Xiaojun Chang, Bingkun Bao, Liqiang Nie, Xinwang Liu, and Yi Yang. The document contains 19 pages and 9 figures and situates itself across the subjects Artificial Intelligence, Computer Vision and Pattern Recognition, and Robotics.

What to watch

Look for follow-up work linked from the paper's associated project URL and for empirical studies that evaluate integrated agents in clinical settings. The paper flags gaps in system-level integration and datasets; progress will be measurable when applied systems demonstrate perception-to-action pipelines validated on representative medical datasets.