To address pain points in science education, including difficulties in representing abstract concepts, the lack of assessment for long-cycle experiments, and fragmented resources between general and special education, we have carried out cross-regional collaborative research together with universities and benchmark basic education schools. This research pioneers a technical architecture of “end-side lightweight design, multimodality, and long-video understanding”. It enables intelligent error correction for long-cycle experiments and multi-dimensional interpretation of abstract concepts at the edge, and develops an AI science teaching tool with wide applicability and low computational requirements. Relying on the "long-term companion Agent" mechanism, we construct a "one-source, multi-track" practical paradigm in real classroom settings. Finally, the multi-track educational reform outcomes are transformed into standardized AI curriculum resource packages, providing a replicable innovation model to boost the high-quality development of inclusive science education in China.

Smart Wearables, especially the ones that integrate wearable e-textiles and machine learning, is an emerging field requiring expertise from textile design, electronic design, machine learning, and human-computer interaction. How to design a course that would teach students relevant skills for smart wearables design and development, as well as transferrable skills required for multidisciplinary collaboration, is an open challenge. This paper presents the design and implementation of a multidisciplinary smart wearables course developed in Aalto University. The course is offered to a mixed cohort that includes students of different levels of study and technical background. This paper shares the experiences from the design and implementation, as well as the challenges of developing inclusive multidisciplinary courses.

Generative AI is in classrooms whether educators planned for it or not; the pedagogical question is no longer whether students will use it but what learning relationship its use creates. We focus on teaching students to build health agents over ubiquitous and wearable sensing data as a stress test for this shift: useful AI participation here rewards students for actively interpreting messy, longitudinal, personally meaningful data and questioning the AI's evidence rather than treating its output as a shortcut. We present Student-AI Collaborative Inquiry (SACI), a pedagogical model that treats the student-AI pair as co-investigators of the student's own multimodal sensor data. SACI rests on three pillars: (1) personal data as the shared object of inquiry, (2) mutual interrogation between student and agent as the central pedagogical mechanism, and (3) a responsible-use scaffold (refusal templates, red-team probes, an IRB decision flowchart) that anchors safety. We illustrate adoption in a short hands-on format and a term-long classroom format, and report design choices, limitations, and an assessment plan for other educators to adapt.

Building, teaching, and maintaining open-source hardware exposes a gap that better documentation does not close: much of what makes a specific machine run in a specific place is tacit, and travels with people rather than files. A repository records the writable surface of a project, its parts, wiring, code, and procedures, but not the work of getting a particular instance running, which stays with the few who built it. Documentation therefore fails at a consistent point, and when it does, a person supplies what was missing; each such moment marks a hole in the infrastructure the documentation could not fill. We report observations from a hands-on open-hardware workshop in which professional scientists, expert in their own fields but new to automation, adopt unfamiliar automation tools, a setting that isolates the problem because a stalled build cannot be blamed on a shaky grasp of the science. We organize these observations as three recurring failures: a setup that runs only for its author, a machine that obeys into a crash, and a machine that runs without working. Capable AI agents do not close these gaps; they make the written-down surface nearly free to produce and leave the tacit remainder exactly where it was. From each failure we draw an implication for how the ubiquitous-computing community should teach and support open hardware.

As artificial intelligence advances, computing education increasingly emphasizes learning process design, yet connecting classroom learning with experiential activities and engineering practice remains challenging. Drawing on practices at Xiamen University’s School of Informatics, this paper proposes a Continuous Learning Experience (CLE) framework integrating Classroom Learning, Experiential Learning, and Authentic Learning. It presents teaching practices involving course organization, hands-on activities, and real-world projects, and examines learning continuity, authentic task-driven learning, and technology-enhanced support. The framework offers practical insights into designing coherent learning experiences and supports the continuous improvement of computing education in response to the opportunities and challenges posed by artificial intelligence.

IoT and UbiComp findings often depend on specific users, devices, environments, and infrastructures, yet conventional paper-reading assignments rarely ask students to consider whether a paper's claims would still be supported if these conditions changed. To make contextual transfer teachable, we present an AI-assisted instructional framework with two linked components. A Trace--Stress--Transfer cycle guides students to connect claims to evidence, identify contextual assumptions, and design resource-bounded transfer tests. A Commit--Challenge--Adjudicate protocol requires students to form an independent judgment before consulting AI and to justify whether its suggestions should be accepted, revised, or rejected. An ExtraSensory walkthrough illustrates the framework, followed by a staged agenda for evaluating its feasibility and educational value.

Data science is vital to modern medicine, but medical curricula rarely include comprehensive training on the topic. The Computing for Medicine (C4M) certificate program at the University of Toronto attempts to address this gap for local medical students and healthcare professionals. Over the course of a 3-part, 15-lecture series, C4M participants progress from writing basic programs to training classical machine learning models on real-world biomedical datasets. This article reflects not only on C4M's syllabus but also on key pedagogical decisions and emerging challenges that have been encountered along the way.

This talk explores integrating mathematical modeling into ubiquitous computing education. It first introduces core modeling concepts, then presents 24 relevant problems selected from three major international mathematical modeling competitions. We analyze these cases across scenarios, data sources, objective functions, constraints, and solution approaches. These curated examples serve as high-quality teaching materials to help researchers apply quantitative methods and educators build students’ problem-solving competencies. The talk will also share reflections on leveraging large language models to assist mathematical modeling practice.