Research
Overview
Our research focuses on identifying and evaluating the appropriate tools required to implement our system, as well as establishing best practices for their operation.
-
Related project review, We examined a variety of successful educational, board, and augmented‑reality (AR) games to understand the factors contributing to their effectiveness and popularity. By analyzing their core mechanics, user engagement strategies, and pedagogical approaches, we aim to incorporate their strongest features into our own design.
-
Technology review, we compared various technology stacks, including game engines, AI models, object recognition models, and image generation models. Our analysis focused on evaluating their performance, compatibility, and ease of integration to determine the most suitable options for our development process.
-
Background research, we explored how gamified learning benefits students by reviewing academic research on game design and reward systems. This helped us design our game logic and mechanics, ensuring they effectively enhance engagement and learning outcomes.
Section 1 - Related Project Review
Section 1.1 - Buzz
One of our project requirements is to implement a “Buzz” game mode. To ensure alignment with our supervisor’s specifications, we analyzed the PlayStation 2 version of Buzz as a benchmark. In Buzz, players participate in a virtual game‑show environment using handheld remotes equipped with four colored answer buttons and a central red “Buzz” button. Questions are presented in two primary formats:
- Speed‑based response: Players attempt to answer as quickly as possible; faster correct responses yield higher point values.
- Buzz‑in format: A question is displayed and players press the “Buzz” button when they recognize the correct answer.
From this review, we concluded that our implementation should feature multiple‑choice quizzes with four answer options to mirror the original design. Furthermore, we will include a dedicated Buzz mode that prioritizes response speed, thereby preserving the competitive, time‑sensitive gameplay characteristic of the original Buzz experience.
Section 1.2 - Pokémon GO
Pokémon GO is an AR‑based mobile game that blends digital gameplay with real‑world exploration, encouraging players to move through their environment to discover virtual creatures. Its success stems from location‑based mechanics and AR overlays that create immersive, contextually relevant interactions.
From this review, we recognized the value of integrating real‑world engagement into our board game by designing activities that prompt players to observe and interact with their surroundings. Additionally, we plan to incorporate optional AR features—such as real‑time, subject‑related 3D meshes generated and overlaid into the environment—to enhance educational content and sustain player motivation through exploratory, interactive gameplay.
Section 1.3 - Mario Party
Mario Party served as the primary inspiration for our board game design due to its widespread popularity among children and its turn‑based, dice‑rolling mechanics, which align closely with our project requirements. In our adaptation, the game board features specialized tiles that trigger unique outcomes when landed on. Rather than traditional mini‑games awarding points by placement, we implemented quiz tiles that present educational questions to players. Performance on these quizzes yields in‑game bonuses that can be strategically applied during subsequent turns, thereby integrating learning objectives directly into core gameplay.
Section 2 - Technology Review
Part of our research is learning about the technologies we have to use as specified by the requirements.
Section 2.1 - Unity Version 6
Our investigation of Unity focused on its Mixed Reality development toolset and cross‑platform deployment capabilities. We evaluated Unity’s support for AR features, performance optimization options, and build pipelines for iOS, Android, and Windows. This analysis informed our decisions regarding engine configuration, platform-specific considerations, and deployment workflows.
Mixed Reality Toolset
Unity has XR Plug‑in Architecture, which is a plugin which can be easily installed and deployed within the editor. It has native support for ARKit (iOS) and ARCore (Android), which is what we need for our Mixed Reality Board Game.
AR Foundation is an extensive framework which can be used to implement Mixed Reality features such as plane detection, image anchors, face tracking, environment probes etc. In particular, its ARPlaneManager and ARRaycastManager can be used to implement the crucial feature of placing the board onto any surface.
A big upside of AR Foundation is its unified abstraction layer, which allows for a single codebase to work for both ARKit + ARCore, and even a non AR version on the desktop. This is very convenient for us as we plan to implement a desktop non-AR version alongside the AR mobile version.
Performance Optimization
Graphics in Unity are handled via three render pipelines:
- Universal Render Pipeline (URP): lightweight, mobile-optimized
- High Definition Render Pipeline (HDRP): high-fidelity, console/PC
- Built-in Pipeline: legacy support
For our project, we chose URP as our render pipeline because it aligns perfectly with our needs. The graphical quality is more than sufficient for our vision, as we are using low-poly models to optimize performance.
Game Engine Comparison
The other more notable game engine out there is Unreal Engine 5*. It is also capable of being used to develop a mixed reality mobile game. We have also researched the features of Unreal Engine 5 and made a table for comparison:
| Feature | Unity | Unreal Engine 5 |
|---|---|---|
| Mixed Reality Support | AR Foundation (unified), mature MRTK | Native AR plugins; experimental MRTK port |
| Render Pipelines | URP (mobile), HDRP (AAA), built-in | Forward Renderer (Mobile), Nanite + Lumen (high-fidelity) |
| Asset & Package Ecosystem | Largest Asset Store; robust third-party support | Growing Marketplace; advanced rendering assets |
| Ease of Use | C# scripting; low entry barrier | C++/Blueprints; steeper learning curve |
As shown above, both game engines have mixed reality support, but only Unity has the unified abstraction layer. Unreal Engine 5 also have a lightweight render pipeline similar to URP so it makes it a viable option.
However, upon browsing on the Unreal Engine 5 asset store, it’s apparent that the Unity Asset Store has a much wider selection. Even for widely used assets such as 2D GUI(Game User Interface) Pack is lacking as only 25 results show up, whereas Unity has thousands. Unity is also easier to develop games in due to it being mostly C# Scripting, which is perfect for our team who have no experience in game development.
In conclusion, Unity is in fact a very suitable game engine for our needs to develop a Mixed Reality Board Game on mobile devices.
Section 2.2 - IBM Granite Model
Our evaluation of the IBM Granite series focused on balancing language understanding performance, memory and compute efficiency, and deployability in offline, performance-constrained environments. We benchmarked multiple versions of Granite models—Granite-2B, Granite-8B and the 4-bit quantized Granite-8B—with a focus on natural language tasks relevant to interactive, such as mixed reality quiz question generation.[1,2,3,4]
Model Performance
We assessed model performance in terms of text generation quality, reasoning ability, and context retention. Granite-2B provided acceptable fluency but lacked depth in contextual reasoning. Granite-8B significantly improved accuracy and coherence in complex tasks. The 4-bit quantized Granite-8B preserved approximately 95% of the full model's performance, with only marginal degradation in language quality and task-specific output.
Memory Usage and Hardware Efficiency
A key consideration was memory footprint and suitability for edge deployment. Full-precision Granite-8B requires approximately 16 GB of VRAM, whereas the 4-bit quantized variant reduces memory usage to 4.5 GB, enabling inference on consumer-grade GPUs and high-end CPUs. Granite-2B is more lightweight (4 GB), but trades off quality.
Online vs Offline Usability
We prioritized models that support offline operation to ensure privacy, stability, and low-latency performance in interactive scenarios. Granite-2B, Granite-8B, and the quantized Granite-8B all support local inference. Among these, the 4-bit model offers the best balance of quality and offline deployability, avoiding reliance on internet connectivity or cloud GPUs.
Conclusion
Based on this analysis, we adopted the 4-bit quantized IBM Granite-8B model for our system. It delivers high-level language performance comparable to full-size models, while enabling efficient, real-time inference in constrained environments. Its compatibility with offline pipelines, low memory footprint, and strong performance across reasoning and generation tasks made it the optimal choice for our mixed reality and mobile deployment needs.
Section 2.3 - YOLO Object Recognition
Our investigation of YOLO focused on its object detection architecture, real-time inference capabilities, and suitability for deployment in mobile and mixed reality environments. We evaluated YOLO’s performance across different versions (YOLOv5, v8 and v11), assessing detection speed, model accuracy, and hardware efficiency. Particular attention was paid to its single-shot detection pipeline, which enables rapid inference suitable for real-time AR interactions. This analysis guided our selection of model and device-specific deployment strategies for performance-constrained environments.
Our selection of the YOLO11x model was driven by its superior detection accuracy and suitability for high-performance deployment scenarios. Among the YOLOv11 family, YOLO11x achieves the highest mean Average Precision (mAP) of 54.7% on the COCO val2017 dataset, significantly outperforming smaller variants.[5]
Despite its increased model size (56.9M parameters) and computational cost (194.9 GFLOPs), YOLO11x maintains acceptable inference latency at 11.3 ms on a T4 GPU and 462.8 ms on CPU. This balance of precision and speed makes it ideal for applications where detection quality is critical, such as in interactive and visually rich environments.
Its strong performance in complex object detection tasks ensures robust real-time behavior when paired with efficient runtime optimization. As such, YOLO11x was selected as the optimal model for our system, where accuracy takes precedence without compromising usability in GPU-enabled settings.
Section 2.4 - DreamShaper Image Generation
Our evaluation of DreamShaper focused on its ability to generate high-fidelity, aesthetically coherent images from natural language prompts, particularly within stylized and imaginative contexts. We assessed its performance in relation to other text-to-image models (such as DALL·E 2, Midjourney, and base Stable Diffusion) by measuring visual consistency, detail richness, and prompt responsiveness. DreamShaper demonstrated superior performance in generating artistically refined images with strong semantic alignment, while maintaining compatibility with standard Stable Diffusion pipelines. Notably, DreamShaper features fast generation speeds and a lightweight model size, making it well-suited for local deployment on resource-constrained devices. This informed our decision to adopt DreamShaper for creative content generation tasks where both visual style and offline usability are prioritized.
Section 3 - Background Research
To create a well-designed game, we conducted comprehensive research into several key areas to understand the non-technical background. Our focus encompassed the integration of games into learning, addressing the needs of students with Special Educational Needs and Disabilities (SEND), and understanding game design principles, particularly reward systems that enhance student engagement.
Integrating Games into Learning
Educational games have been shown to significantly enhance student engagement and learning outcomes. Research indicates that gamification can increase motivation, promote active learning, and improve retention rates among students. For instance, a study from the University of Warwick found that incorporating games into teaching improved student exam performance, with the median student achieving a 69% grade—just shy of a first-class mark—compared to 60% in traditional teaching methods[1]. Further research indicates that gamification can increase motivation, promote active learning, and improve retention rates among students[2]. By incorporating game elements such as challenges, feedback, and rewards, educational content becomes more interactive and appealing, leading to a more profound learning experience.
Addressing the Needs of Students with SEND
Students with SEND often face unique challenges in traditional educational settings, including difficulties with concentration, social interaction, and sensory processing. Gamification has been identified as a beneficial tool for supporting these students[2]. Games often provide structured environments with clear rules and immediate feedback, which can help students with SEND, particularly those with autism, to focus and engage more effectively[3]. For instance, visual schedules and predictable patterns in games can reduce anxiety and improve task completion rates among autistic learners.
Designing Reward Systems in Games
An effective reward system is crucial for maintaining player engagement and motivation in educational games. A study published in the International Journal of Human-Computer Studies examined how the placement of rewards within an application affects the frequency of its use[4]. The findings suggest that users are more likely to engage with an application when rewards are presented early in the interaction. This occurs due to temporal discounting, where rewards lose value the longer they are delayed. Therefore, implementing rewards as close as possible to the start of an interaction can enhance their effectiveness in motivating users.
When designing incentive systems, it's important to balance external rewards (like points and badges) with internal motivation (personal enjoyment). Relying too much on external rewards can reduce a person's natural interest in an activity—a phenomenon known as the overjustification effect. This occurs when external incentives overshadow the inherent enjoyment of an activity, leading individuals to attribute their participation to the reward rather than personal interest[5]. Therefore, effective reward systems should not only offer external incentives but also enhance the intrinsic satisfaction derived from the learning experience.
Additionally, research involving primary school children interacting with digital game-based learning environments indicates that while students may report excitement about game incentives[6], their visual attention to these elements during gameplay is minimal. This suggests that while reward systems are important, their design should not overshadow the educational content but rather complement it to maintain a balanced and effective learning experience.
In summary, our background research underscores the importance of thoughtfully integrating game elements into educational contexts, particularly for students with SEND. By designing reward systems that consider the timing and nature of incentives, we aim to create an engaging and supportive learning environment that caters to the diverse needs of all students.
Section 4 - Technical Decisions
After extensive research, we finalized the following technological stack for our AR board game:
| Type | Decision |
|---|---|
| Game | Unity 6 engine |
| AR generation | Unity AR Foundation 6.1.0 |
| AI Model | 4-bit quantized version of Granite-3.0-8B instruct model |
| Object recognition | OpenCV YOLO11x |
| Image generation | DreamShaper-8 |
| Teacher Dashboard | React+Vite website |
These technologies were selected based on their ability to support real-time AR interactions, AI-driven adaptability, and user-friendly interfaces for both students and educators. The combination of these tools ensures a robust, scalable, and effective learning platform tailored to SEND students' needs.
Section 5 - References
[1] ibm-granite, “granite-3.0-language-models/paper.pdf at main · ibm-granite/granite-3.0-language-models,” GitHub, 2024. https://github.com/ibm-granite/granite-3.0-language-models/blob/main/paper.pdf (accessed Mar. 27, 2025).
[2] “ibm-granite/granite-3.0-8b-instruct · Hugging Face,” Huggingface.co, Feb. 24, 2025. https://huggingface.co/ibm-granite/granite-3.0-8b-instruct (accessed Mar. 27, 2025). [3] “QuantFactory/granite-3.0-8b-instruct-GGUF · Hugging Face,” Huggingface.co, Oct. 21, 2024. https://huggingface.co/QuantFactory/granite-3.0-8b-instruct-GGUF (accessed Mar. 27, 2025).
[4] Huggingface.co, 2025. https://huggingface.co/QuantFactory/granite-3.0-8b-instruct-GGUF?show_file_info=granite-3.0-8b-instruct.Q4_K_S.gguf (accessed Mar. 27, 2025).
[5] “Ultralytics/YOLO11 · Hugging Face,” Huggingface.co, 2017. https://huggingface.co/Ultralytics/YOLO11
[6] Ross, J. et al. (2024) Using games in teaching ‘boosts grades and student satisfaction’, Times Higher Education (THE). Available at: https://www.timeshighereducation.com/news/using-games-teaching-boosts-grades-and-student-satisfaction#:~:text=The%20study%20found%20that%20the,cent%20in%20the%20control%20group. (Accessed: 27 March 2025).
[7] Lisa-Maria Putz. (2020) Can gamification help to improve education? findings from a longitudinal study, Computers in Human Behavior. Available at: https://www.sciencedirect.com/science/article/abs/pii/S074756322030145X (Accessed: 27 March 2025).
[8] Hickman, C. (2023) How gamification can support autistic learners, Organization for Autism Research. Available at: https://researchautism.org/blog/how-gamification-can-support-autistic-learners/ (Accessed: 27 March 2025).
[9] Atherton, G. and Cross, L. (2021) The use of analog and digital games for autism interventions, Frontiers in psychology. Available at: https://pmc.ncbi.nlm.nih.gov/articles/PMC8384560/ (Accessed: 27 March 2025).
[10] Garaialde, D. (2021) Designing gamified rewards to encourage repeated app selection: Effect of reward placement , Designing gamified rewards to encourage repeated app selection: Effect of reward placement. Available at: https://discovery.ucl.ac.uk/id/eprint/10131140/1/1-s2.0-S1071581921000793-main.pdf (Accessed: 27 March 2025).
[11] Kendra Cherry, Mse. (2023) Why does the overjustification effect reduce intrinsic motivation?, Verywell Mind. Available at: https://www.verywellmind.com/what-is-the-overjustification-effect-2795386 (Accessed: 27 March 2025).
[12] Li, Y., Chen, D. and Deng, X. (2024) The impact of digital educational games on student’s motivation for learning: The mediating effect of learning engagement and the moderating effect of the digital environment, PloS one. Available at: https://pmc.ncbi.nlm.nih.gov/articles/PMC10783726/ (Accessed: 27 March 2025).