Thesis Description & Objectives
Virtual Reality (VR) has proven effective in simulating procedural tasks and structured training scenarios, particularly in domains where experiential learning is critical. However, current VR learning experiences are often limited in their ability to support dynamic problem-solving, especially in situations where learners are required to devise tools, strategies, or solutions in real time. This lack of adaptability reduces learner agency and constrains the exploration of alternative approaches within immersive environments.
This thesis aims to investigate the use of Generative Artificial Intelligence to enhance learner agency in VR-based training systems. The objective is to enable users to interactively solve training scenarios by leveraging AI to generate context-aware suggestions, including tools, hypotheses, and procedural steps. In addition, the work will explore the integration of Reinforcement Learning techniques to adapt the behavior of the AI assistant over time, optimizing when and how support is provided based on user interactions and learning progress.
The student will design and develop prototype systems in which users can invoke AI assistance during VR training sessions. The approach will initially focus on unconstrained environments, where learners can freely explore AI-generated solutions in a sandbox setting. Subsequently, domain-specific constraints will be introduced to guide the generation process toward feasible and coherent solutions aligned with the procedural logic of the task. Reinforcement Learning strategies will be investigated to model adaptive assistance policies, allowing the system to personalize feedback, regulate the level of guidance, and improve its effectiveness through interaction. Particular attention will be devoted to balancing flexibility and control, as well as to the definition of meaningful reward signals based on user behavior, task completion, and learning outcomes.
The developed systems will be evaluated through qualitative and quantitative methods, considering aspects such as user engagement, perceived agency, solution diversity, coherence with domain knowledge, and overall usability. The expected outcome is an enhancement of learner creativity and autonomy through AI-assisted problem solving, along with the development of adaptive learning pathways tailored to user behavior. This work may also contribute to the definition of intelligent, responsive training environments capable of evolving in real time through the combined use of Generative AI and Reinforcement Learning.
References
[1] S. Doroudi, V. Aleven, and E. Brunskill, “Where’s the Reward? A Review of Reinforcement Learning for Instructional Sequencing,” in International Journal of Artificial Intelligence in Education 29, 2019, pp. 568-620. doi: 10.1007/s40593-019-00187-x.
[2] S. Yao et al. “ReAct: Synergizing Reasoning and Acting in Language Models,” arXiv:1907.00456, 2019, pp. 1-16, doi: 10.48550/arXiv.1907.00456.
[2] N. Jaques et al. “Way Off-Policy Batch Deep Reinforcement Learning of Implicit Human Preferences in Dialog,” arXiv:2210.03629, 2023, pp. 1-33, doi: 10.48550/arXiv.2210.03629.
