Throughout our lifetime, we are acquiring and improving a wide range of motor skills. In other words, we are able to perform certain movements faster, better and more accurately. The mechanisms at play in motor skill acquisition are complex and not yet fully understood, but there is a consensus that practice is fundamental to skill learning (Ericsson et al., 1993). For instance, a novice musician would spend hours practicing her instrument in order to play music pieces, or to play with other musicians. The same is true for athletes. Practice plays a fundamental role in motor learning and it can, ideally, be designed in order to improve learning performance. However, practice design is challenging as it relies mostly on heuristics (from the learner, a teacher, or a pedagogy), tacit knowledge and varies across learners and along the learning development. Designing good practice sessions is all the more important for beginners who do not have the expertise to select efficient practice routines and to understand what makes a good or a bad practice schedule.
In the ARCOL project we envision computational solutions able to learn novice individual-specific practice design in order to support the acquisition of motor skills.
We foresee a technological solution able to capture data representing movement execution (e.g. wearing sensors), able to learn about mechanisms of skill acquisition from data streams (machine learning), and, reciprocally, able to guide a human in the process of skill acquisition (e.g. through instructions or demonstrations). We consider two use cases: motor learning in piano performance, and motor learning in prosthesis control (these use cases are further detailed below).
At the core of the project lays the use of machine learning as a means to understand and facilitate motor learning. We call co-learning the resulting dual process through which both human and machine learn reciprocally. Both types of machine and human learning mechanisms have been discussed at a theoretical level in the literature under the formalism of reinforcement learning, which will be used in this project. In reinforcement learning, a system learns to take an action (e.g. choosing an instruction like a motor task) according to the expected reward from having taken this action. Facilitating skill acquisition through reinforcement-driven learning systems remains however unexplored. The project is fundamentally multidisciplinary and contributes to the fields of Human- Machine Interaction and Motor Learning.