本論文針對高中生提出一個深度學習視覺分類機器人平台與一套基於資料驅動學習的人工智慧機器人課程,讓其可以在高中職階段實施和推廣人工智慧機器人的基礎教育。主要有三個部分:(1) 深度學習視覺分類機器人平台、(2) 人工智慧機器人課程、以及(3) 資料驅動學習。在深度學習視覺分類機器人平台方面,本論文使用低價位的邊緣運算裝置來設計一台具深度學習能力之小型移動機器人,使其具有價格合理、軟硬體彈性高、場地模組化、以及擴充性高的特色。此外,搭配本論文提出之輕量型卷積神經網路模型,使其可在一般規格之電腦上有理想的訓練速度,且訓練後的神經網絡在邊緣運算裝置上也有不錯的推論準確度與執行速度。在人工智慧機器人課程方面,本論文設計一套24小時動手做課程,包含了人工智慧觀念、影像處理演算法、深度學習神經網路、以及機器人控制等四個單元,使其具有概念學習、動手操作、以及錯誤釐清的特色。在資料驅動學習方面,本論文設計一個自駕車的場地與情境。在路牌辨識之實際操作過程中,讓學生可以瞭解所蒐集之照片資料的品質與數量對於神經網路在學習上的影響。從分析與訪談的結果可知,本論文所提出之平台與課程確實可以在高中職教學現場成功地實施,並且可以提高學生的學習效果以及建立正確之人工智慧與機器人控制的概念。
In the dissertation, a deep learning visual classification robotics platform and an artificial intelligence robotics curriculum based on data-driven learning are proposed for high school students so that the basic education of artificial intelligence robots can be implemented and promoted in high school. There are three main parts: (1) deep learning visual classification robotics platform, (2) artificial intelligence robotics curriculum, and (3) data driven learning. In deep learning visual classification robotics platform, some low-cost edge computing devices (Raspberry Pi SBC and Intel NCS compute stick) are used to design a small-sized mobile robot with deep learning capability so that it has the characteristics of reasonable price, high flexibility of hardware and software, modularity of venue, and high expandability of functions. In addition, a lightweight convolutional neural network model is proposed so that this network model can be fast trained on ordinary computers and the trained neural network can also have a pretty good inference accuracy and execution speed on this edge computing device with limited performance. In artificial intelligence robotics curriculum, a 24-hour hands-on curriculum is designed. It includes four units: artificial intelligence concepts, image processing algorithms, deep learning neural networks, and robot control so that it has the characteristics of concept learning, hands-on operation, and misunderstanding clarification. In data driven learning, a venue and a scenario of a self-driving car are designed. During the actual operation of neural network training for the recognition of road signs, students can understand that the quality and quantity of the collected photo data have a great influence on the learning of neural networks. From the results of analysis and interviews, it can see that the platform and the curriculum proposed in this dissertation can be successfully implemented in the teaching site of high school, and can improve the learning effect of students and establish the correct concept of artificial intelligence and robot control.