Abstract

Human rowing coaches, who are experienced in describing how athletes move, could deputise machines to provide similar feedback if they could teach them what to look for. Ubiquitous and precise kinetic sensors, along with probabilistic inference algorithms, have strengthened the foundations of motor-skill based interactive systems. However, these systems still leave us uncertain over how to feel about a particular performance.

I will discuss a framework for developing machines that rate athletes along natural and emotive scales such as “the importance of improving how much `suspension’ they achieve”. After presenting a model of the criteria rowing coaches use to describe their judgements of indoor rowing techniques, I will propose an iterative scheme for single criteria that resolves multiple judges’ perspectives on a set of performances into a consensus of ratings for each performance. I will share guidelines for collecting data on indoor rowing techniques for machine learning, before presenting initial evidence in support of training Bayesian models to predict human ratings. I will evaluate generative filters and linear regression for four criteria, showing they give more useful predictions than 1) 10% to 50% of the human coaches, 2) randomly rating performances (1×10-6 < p < 0.38) and 3) performance indicators based on Newtonian mechanics (1×10-3 < p < 0.6).