"AlphAI": Teaching AI Algorithms to K12 by Training Learning Robots and Visualizing How It Works

Marie Absalon, Thomas DeneuxAAAI 2025 · DOI

Abstract

AlphAI is an educational ecosystem designed to demystify artificial intelligence for K-12 students. By integrating physical robotics with real-time visualization of internal algorithmic processes, AlphAI moves beyond "black box" interfaces. The system employs a four-level pedagogical framework—ranging from manual control to reinforcement learning—to foster AI literacy and metacognition. Experimental results show significant improvements in AI knowledge and student perseverance, demonstrating the effectiveness of tangible, transparent AI education.

Introduction

As AI becomes ubiquitous, "AI literacy" is transitioning from an elective to a core requirement in K-12 education. However, current pedagogical tools often fall into two traps: they are either overly abstract, focusing on high-level mathematics and code, or they function as "black boxes." Tools like Teachable Machine allow students to provide data and receive results without understanding the underlying mechanisms. Furthermore, while Supervised Learning is common in classrooms, Reinforcement Learning (RL)—central to modern autonomous systems—is rarely taught to younger students. AlphAI addresses these gaps by making the internal "thinking" of an AI visible and tangible through robotics.

The AlphAI Ecosystem

The authors developed AlphAI, a software-hardware resource that integrates with popular educational robots such as Lego Spike, mBot, and Thymio. The methodology relies on three pillars:

  1. Tangible Interaction: Students train robots for physical tasks like circuit racing or obstacle avoidance, making AI behavior observable.
  2. Real-Time Visualization: AlphAI "opens the box" by visualizing neural network weights (via line thickness and color), K-Nearest Neighbor (KNN) decision boundaries, and Q-learning values as they update.
  3. The "4 Levels of Autonomy" Curriculum: A framework guiding students from manual remote control (Level 1) and hard-coded programming (Level 2) to Supervised Learning (Level 3) and Reinforcement Learning (Level 4).
The AlphAI interface layout, showing the mapping from sensor input to internal algorithmic details and resulting robot actions.

The system emphasizes user-generated datasets. Students collect training data using robot sensors (infrared, ultrasonic, or low-resolution camera pixels). For example, a "2-pixel camera navigation" dataset helps students visualize 2D decision boundaries and understand concepts like overfitting and nonlinearity.

Pedagogical Insights

By using simple sensors instead of high-resolution video, the internal mathematics of the AI becomes simple enough to visualize completely. Students can see exactly which sensor triggers which "neuron." This transparency allows for "Metacognitive Mirroring," where teaching a robot to learn through trial and error helps students reflect on their own learning processes, reinforcing the idea that mistakes are necessary for optimization.

Visualization of KNN decision boundaries and neural network architectures within the AlphAI software.

Furthermore, the software allows students to "program" a neural network manually by drawing connections before letting the algorithm take over. This demystifies the transition from classical logic to machine learning. Because students collect their own data, they also encounter "implicit biases" firsthand (e.g., a robot only turning left because it was only trained on left turns), providing a concrete basis for discussing AI ethics.

Evaluation and Results

The AlphAI ecosystem was tested in elementary schools and university settings:

  • Elementary (Grades 3-5): In a study of 138 students, AI knowledge scores increased from 39% to 65% following a 6-hour workshop.
  • Metacognition: Student agreement with the importance of perseverance during difficult tasks rose from 65% to 81%.
  • University Level: The curriculum has been integrated into an "Initiation to AI" course for non-computer science majors at Paris-Saclay University, reaching 250 students annually over three years.

Discussion and Future Work

While effective for classical machine learning and "toy models," a gap remains between these small-scale models and the massive Generative AI (LLMs) students use daily. Future work aims to develop "toy models" of generative architectures to bridge this gap. Additionally, mapping the curriculum to diverse national education standards remains an ongoing task.

Summary updated

BibTeX
@inproceedings{absalon2025alphai,
  author = {Marie Absalon and Thomas Deneux},
  title = {{"AlphAI": Teaching AI Algorithms to K12 by Training Learning Robots and Visualizing How It Works}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35181},
}
Figure