Agen­t­ic Vir­tu­Tutor: Im­ple­ment­a­tion of an agen­t­ic em­bod­ied AI learn­ing as­sist­ant and com­par­is­on with a non-agen­t­ic im­ple­ment­a­tion in the ser­i­ous VR game Vir­tuChem­Lab

With the growing importance and capabilities of serious games and generative AI (GenAI), the VirtuChemLab has been fitted with a GenAI-powered virtual laboratory tutor called VirtuTutor. The role of this tutor within the simulated chemistry laboratory is to help users resolve difficulties, answer their questions, and improve learning outcomes without human supervision. 

In my thesis, I replaced VirtuTutor’s basic GenAI components with an agentic AI system to address several limitations of the original components. I evaluated how the agentic implementation improves response quality across 10 attributes, usability including  cognitive load, as well as the amount of chemical misinformation, in comparison to the non-agentic version. To contextualize these results, I also evaluated the consumption of computing resources. I found that the agentic VirtuTutor significantly improves response quality, usability including cognitive load, while also reducing the amount of chemical misinformation. At the same time, the agentic VirtuTutor consumes significantly more computing resources. 

In my presentation, I will outline the motivation and research questions behind this work. Then, I will explain the terminology surrounding agentic AI and present my final agentic implementation. Finally, I will outline my evaluation design and present the results of these evaluations, providing answers to my research questions. I will also propose future work for the agentic VirtuTutor.