Embodied AI systems that act in the physical world carry a distinctive risk: digital failures and attacks can cause physical harm. A new survey on arXiv takes this cyber-physical risk as its starting point and organizes the field's safety concerns into a structured framework.

The survey organizes embodied AI safety through a four-layer reference architecture, according to the abstract. The paper's title, 'From bits to atoms,' signals that the scope runs from digital computation to physical actuation. As a survey, its contribution is a cross-layer map of existing work rather than a new safety method.

The significance of the framing is that safety is treated as a property of the whole system, not of individual components. Because a failure at any layer can propagate into the physical world, the survey points to the need for safety measures that span all layers.