Embodied agents operating in complex, open-ended environments require autonomy beyond predefined tasks and human guidance. Traditional systems, though effective under controlled conditions, often fail when facing uncertainty, ambiguity, dynamic environments, or evolving goals. To achieve autonomous learning like a baby, autonomous embodied Artificial Intelligence (AI) integrates three core capacities: Autonomous Environment Learning, which enables agents to autonomously explore, perceive, and interpret the dynamic and open environment; Autonomous Cognitive Learning, which focuses on the autonomous development of reasoning and decision-making skills; and Autonomous Dynamic Interaction, which facilitates agents grounded in environmental learning and cognitive models to engage in autonomous dynamic interaction with entities (who/what) and at the right time (when). These capacities are not independent modules but interconnected processes forming a continuous loop of perception, cognition, and interaction. Together, they allow agents to reduce uncertainty, expand action possibilities, and generate meaningful behavior without external instruction. This framework provides a foundation for advancing embodied AI in domains such as robotics, autonomous driving, and scientific discovery, where long-term adaptability and independent decision-making are essential.