Dynamic Preference Learning for Complex Cobotic Tasks

In complex tasks requiring robots to collaborate with multiple human users for successful human-robot interaction, it is essential to capture user preferences in the control design. This work considers both the subjective judgements of a users preferences, to formulate a preference measure that can be used in control design. Novel to the work, is the consideration of the temporal impact that current preferences have on previously provided preferences. We formulate a weighted set-wise preference learning problem that considers the historical preferences of an individual, capturing the impact of memory on human user's preferences. We address the challenge of how to choose the best preference features as the individuals set-wise preference comparisons are sequentially presented by designing a weighted sequential preference estimator. We build upon the concept of an individual's preferences to a group of users' preferences, modelling group preferences as a normal distribution. We further extend the sequential individual preference estimation approach to multiple users' preference through the derivation of a stochastic gradient descent group preference estimator. An experimental study with 43 participants is conducted to support the proposed approaches. Copyright (c) 2023 The Authors.