A Survey of Agentic Robotics: Toward Continual Self-Improvement

Foundation models (FMs), particularly large language models (LLMs) and vision-language models (VLMs), are increasingly used as the decision-making core of AI agents and are now being extended from digital environments to robotic systems. In classical robotics, engineers decompose tasks, specify objectives, write control code, and repair failures before deployment; in agentic robotics, an FM makes some of these decisions at run time or during improvement. Bringing this paradigm into the physical world introduces additional challenges: FM-mediated decisions must be grounded in executable robot behavior, respect physical constraints, and close the loop through real-world feedback. We survey this emerging paradigm and organize existing systems by three loci of FM-mediated agency: task orchestration, motion/action generation, and improvement. Coding rules for each locus make the definition operational, including a test of which improvement decisions the FM makes that separates fixed-procedure from agent-directed improvement. We apply these rules to the systems retrieved by a systematic search, examine verification and feedback-closure mechanisms, and compare what self-improving systems have demonstrated. Finally, we discuss continual self-improvement as a central emerging frontier and the challenges toward reliable autonomous improvement through physical interaction.