New framework defines trustworthiness for embodied AI with graded levels

A new paper from 40+ researchers at multiple institutions presents a comprehensive framework for trustworthy embodied intelligence, addressing the gap between task completion and actual trustworthiness in systems that integrate perception, decision-making, and physical interaction. The authors define trustworthiness as the sustained capacity to execute specified tasks reliably while keeping risk within acceptable bounds—termed 'sustained safe success.' The framework organizes support into four interdependent layers: the model layer generates action proposals with calibrated uncertainty and safety preferences; the system layer ensures dependable execution via sensing, computation, control, and hardware safeguards; the evidence layer substantiates claims through evaluation, verification, and traceability; and the deployment layer maintains validity via runtime monitoring, incident response, and controlled updates. Because failures propagate across layers, no single component can guarantee end-to-end trustworthiness.

Building on insights from embodied AI, robotics, control theory, and autonomous driving, the authors propose a non-normative hierarchy of trustworthiness levels that grades the strength of bounded deployment claims. This hierarchy spans task capability, safety, system assurance, operational governance, and supporting evidence. The goal is to provide a basis for comparative evaluation, research prioritization, and future standardization. By framing trustworthiness as a multi-layered, graded property, the work offers a structured path for deploying embodied systems—such as autonomous robots and vehicles—in real-world environments where failures can cause immediate physical harm. The paper is available on arXiv with code and additional resources linked.