E2HRA: A Safety-Governed End-to-End Reference Architecture for General-Purpose Humanoid Robots

A humanoid robot is not simply an anthropomorphic mechanical platform coupled to a large artificial intelligence model. It is a cyber-physical system that must continuously reconcile embodiment, energy, sensing, contact dynamics, cognition, real-time action, human intention, and physical safety. Contemporary research offers increasingly capable foundation models, whole-body controllers, and simulation environments; however, an implementation-independent description of how these elements are integrated, governed, and verified remains essential. This paper introduces E2HRA (End-to-End Humanoid Reference Architecture), a proposed end-to-end systems-engineering reference architecture for general-purpose humanoids. E2HRA defines eight operational planes, from embodied mechanics and embedded interfaces through real-time control, locomotion and manipulation, skill supervision, multimodal world models, deliberative cognition, and human/cloud services. Three cross-cutting assurance domains—independent functional safety, cybersecurity and data governance, and digital-twin-based verification—constrain the complete system. The principal architectural contribution is not a new trained robot policy, but a set of explicit subsystem boundaries, typed interface contracts, timing and fault semantics, a closed-loop execution protocol, and an evidence-based deployment process. A worked household-assistance scenario and a proposed test matrix illustrate how a task can be translated into guarded action and how failures are handled. A low-cost simulation-first implementation path is provided alongside optional high-performance extensions.