Agentic Embodied AI for Human–Robot Collaboration: Architectures, Technologies, and Future Directions

Human-robot collaboration has relied on task models that are specified in advance for each product. Embodied foundation models and agentic artificial intelligence change this premise, because a robot can now interpret an open goal, plan, act, monitor the outcome and recover from failure. This review examines what that change means for collaboration. A critical narrative review was conducted over 231 sources. Embodied intelligence, agentic intelligence, autonomy and collaboration are distinguished, and four collaboration modes are defined by the roles of the human and the robot. Four system architectures are compared: modular agentic, end-to-end embodied, hierarchical agentic embodied, and human-centred shared autonomy. Six groups of enabling technologies and five collaboration mechanisms, from autonomy allocation to failure recovery, are analysed. Applications in six domains, from manufacturing to domestic service, are assessed by the maturity of their evidence. Reported systems are validated mainly on single cells and short tasks, human factors are measured in few studies, and verified safety layers remain separate from learned reasoning. An evaluation framework and research directions for seven open challenges are proposed. The review gives researchers and system builders a basis for bringing embodied and agentic AI into human-robot collaboration and for evaluating the systems that result.