Researchers from the UK demonstrate the efficacy of self-supervised learning for human activity recognition using a vast wearable dataset. By addressing convergence issues and evaluating multi-task self-supervision, they showcase robust models with superior representation quality across diverse populations and activity types.