Reimagining Machine Learning's Role in Assistive Technology by Co-Designing Exergames with Children Using a Participatory Machine Learning Design Probe
Jared Duval, Laia Turmo Vidal, Elena Márquez Segura, Yinchu Li, Annika Waern · 2023 · Proceedings of the 25th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2023) · doi:10.1145/3597638.3608421
Summary
This paper fundamentally reframes the role of machine learning in assistive technology, arguing that ML models do not need to be accurate to be valuable — they can serve as sources of play and motivation rather than diagnostic tools. The researchers developed Cirkus, a smartphone-based design probe that supports animal locomotion exergames where children move like various animals (kangaroos, frogs, bears, lizards, etc.) while the app collects accelerometer, rotation, and bearing data from wrist-worn devices. The work is situated within a collaboration with Cirkus Cirkör, the largest circus group in Sweden, developing playful physical therapy for children with Sensory Based Motor Disorder (SBMD) — a condition within Sensory Processing Disorder that affects how the brain integrates sensory, movement, and positional information, impacting balance, coordination, and motor planning. SBMD symptoms are present in around 16% of the population and frequently co-occur with ADHD and autism. Circus arts were chosen as the therapeutic framework because they promote balance, coordination, and social skills through cooperative rather than competitive activity. The design process began with bodystorming sessions with high school interns, followed by iterative development of a generalized play framework built around three configurable settings: who chooses the animal, who performs, and who rates. Five participatory workshops were conducted with a total of 30 children aged 10-12, including children with and without SBMD, across circus facilities, a school, a university lab, and a dedicated SBMD co-creation session. The workshops produced a catalog of 17 exergames and 673 instances of time series movement data totaling 569 megabytes.
Key findings
The research produced three key contributions. First, the Cirkus design probe successfully supported participatory game co-creation while simultaneously collecting movement data for ML training — demonstrating that play naturally affords the repetitive movement that ML requires for training data, without the process feeling tedious. Children were enthusiastic about "teaching" the technology, and the games they co-created (including Labyrinth, Hunter Gatherer, Animal Crossing, Guess, Popcorn, and Rise!) showed remarkable variety while all being supportable by the flexible three-setting framework. Second, the resulting ML model — a 1.1 megabyte neural network trained on the collected movement data — was intentionally inaccurate, but the authors argue this inaccuracy is a feature, not a bug. Rather than using ML to diagnose or assess children's movement quality (which risks encoding normative assumptions about "correct" bodies), the inaccurate model is positioned as a "classification monster" or playful antagonist that creates motivating friction in games, similar to "grinding" mechanics in commercial games. Third, circus professionals requested removing the competitive rating feature because it conflicted with circus's collaborative ethos, leading to a "no one rates" option — a finding that highlights how stakeholder context directly shapes what participatory ML looks like in practice. One child with co-occurring autism who was becoming overstimulated was given the facilitator role (controlling the iPad), which calmed them and enabled continued participation.
Relevance
This paper makes a provocative and important contribution to how the accessibility field thinks about machine learning. The dominant narrative positions ML accuracy as paramount, but this work demonstrates that in therapeutic and playful contexts, inaccuracy can be intentionally leveraged for motivation and engagement. This challenges three assumptions common in assistive technology: that ML must be accurate to be useful, that data collection for ML must be separate from the application itself, and that ML training requires expert-controlled processes. For practitioners working with children with disabilities, the Cirkus framework offers a concrete model for how technology can support physical therapy through play without medicalizing the experience. The finding that children were motivated to "teach" the technology — giving them agency over the ML pipeline rather than being subjects of it — connects to broader disability justice principles about who controls assistive technology. The work also provides practical guidance on collecting movement data from children with disabilities in participatory contexts, noting that play naturally generates the repetitive data ML needs without requiring tedious, clinical data collection sessions.
Tags: machine learning · participatory design · exergames · children · sensory processing · physical therapy · co-design · game accessibility · assistive technology · design probe