Embodied AI: Bridging Simulation and Reality in Robotics

This paper presents a comprehensive review of simulation-to-reality (Sim2Real) transfer techniques in the context of embodied artificial intelligence (embodied AI). Embodied AI refers to artificial intelligence systems that interact with the environment through a physical or virtual body (e.g., robots, virtual agents), enabling perception, decision-making, and action in real or simulated spaces. We focus on their applications in robotic learning and control. By critically comparing the capabilities of mainstream simulation platforms such as Habitat and Isaac Gym, the study identifies core trade-offs between physical realism and computational efficiency. It further analyzes transfer methodologies, including domain randomization, domain adaptation, and hybrid techniques, highlighting their effectiveness and limitations in dynamic and unstructured environments. The review also investigates the emerging role of large multimodal models, such as PaLM-E and RT-2, in bridging semantic understanding with robotic action. While these models offer significant improvements in generalization and task planning, challenges remain in achieving real-time performance and physical grounding. The findings suggest that simulation fidelity alone does not guarantee successful transfer, hybrid transfer methods outperform single-strategy approaches in complex settings, and large language models hold promise for enhancing robot intelligence but must be optimized for embedded deployment. This study offers practical insights for developing scalable and cost-effective Sim2Real pipelines in industrial applications such as autonomous navigation and robotic manipulation. It further outlines future directions in edge computing, lightweight modeling, and ethics-aware simulation, promoting the integration of physics-aware AI in real-world robotic systems.