A Systematic Approach to Define Realism in Embodied AI

Introduction: Robots are showing up more in everyday life, and people are interacting with them more often. They are being used in various fields, such as healthcare and education etc. The growing presence of robots shows clear progress in how robots move, act on their own, and interact intelligently. However, there has been little focus on realism in embodied AI. Realism in robots encompasses much more than superficial appearance. It can include safe and reliable interaction with people, social or behavioral realism like trustworthiness, human‑like adaptability in dynamic environments. The realism of a robot plays an important role in shaping users’ experiences. If a robot looks similar to a human but acts in an unexpected way, users face an uncanny experience or the Uncanny Valley effect. If robots look artificial or mechanical, it is difficult for people to have a meaningful interaction. Making a comprehensive definition and right balance in a robot’s realism is essential to ensure positive user experiences and appropriate interaction. Methods: The methodology used in this study to explore realism in Embodied AI was the Systematic Literature Review. Firstly, clear research questions were defined, followed by the development of inclusion and exclusion criteria. Then a comprehensive search was done in electronic databases to identify the relevant papers. Different keywords and filters were applied to find the appropriate studies. A PRISMA flow diagram was used to illustrate the steps taken to select studies and papers for the systematic review. In the end, 110 papers were selected for further screening. After reviewing the title, abstract, and full text of the papers, 48 papers were chosen for analysis regarding realism in embodied AI. This approach helped select the studies that were most relevant and useful for my research. Results: Fourteen dimensions have been identified in this study that reflect realism in embodied AI 1)humanlike or anthropomorphism 2)eye gaze 3)gesture realism 4)natural interaction 5)voice and speech naturalness 6)response timing 7)intuitive Interaction 8)real and believability 9)visual and physical embodiment 10)predictability 11)perceived Intelligence 12)empathy and social norms adherence 13)adaptivity, 14)privacy and data respect. The most significant for robot realism were natural interaction, real and believability and perceived intelligence. It shows that most of studies emphasize how naturally a robot interacts with humans and how believable and smart the robot appears in users' perspective. These dimensions were categorized into five clusters 1)Interactive realism, 2)physical/visual realism, 3)behavioral realism, 4)ethical/moral realism, 5)cognitive/social realism which helped to make a hierarchical taxonomy. Conclusion: A comprehensive understanding of how realism was conceptualized and assessed in embodied AI has been developed in this thesis. We found that realism in embodied AI was multilayered and interconnected. Among the dimensions, natural interaction and real and believability emerged as the most influential dimensions due to their strong presence across the literature, while perceived intelligence stood out for bridging multiple layers of realism, and it was presented in several clusters more than other dimensions. Based on these main dimensions, a realistic robot interacts easily and naturally with people in regular social situations, responds to context in an appropriate way and expresses behaviors that align with social standards.