The Unexpected Daily Situations (UDS) Dataset: A New Benchmark for Socially-Aware Assistive Robots

This article presents the progress in building a new dataset of 'unexpected daily situations' (like someone tripping on a box, while carrying a tray to the kitchen, or someone burning him/herself with hot water and dropping a mug). Each of the situations involve one or two humans in a familiar, structured environment (eg, a kitchen, a living room) with rich semantics. Correctly interpreting the situation (including recognising an error, undesired effect or incongruity when it occurs, as well as selecting the best repair action) requires beyond-state-of-art spatio-temporal, semantic and socio-cognitive modelling. As such, the aim of the dataset is to offer (i) a realistic source of data to train and test such novel algorithms and (ii) provide a new benchmark against which algorithms can be demonstrated.