Multimodal Uncertainty Reduction for Intention Recognition in a Human-Robot Environment

Author: Susanne Trick

Supervisors: Prof. Constantin Rothkopf, Ph.D.; Prof. Jan Peters, Ph.D.; Dorothea Koert

Submission: November 2018

Abstract:

In order to improve the quality of life and personal independence of an increasing amount of elderly people, assistive robots can potentially be applied for supporting everyday life activities. To guarantee a safe and intuitive interaction between human and robot the human intentions need to be recognized automatically. The use of multiple modalities for this intention recognition may not just increase the robustness against failure of individual modalities but can especially be used to reduce the uncertainty about the intention to be predicted.

Thus, in contrast to existing methods, in this work, an approach for multimodal intention recognition is introduced that focuses on uncertainty reduction through Bayesian classifier fusion. For this, a Hierarchical Bayesian model is applied through which classifiers can be fused that output Categorical probability distributions. Meanwhile, both their individual uncertainty and reliability in terms of accuracy determine the individual distributions' impact on the fused result distribution. Since the proposed fusion approach should be applied for multimodal intention recognition, additionally individual intention classifiers for the four modalities speech, gestures, gaze directions and scene objects were trained. Each of them outputs a probability distribution over all possible intentions. The distributions returned by the individual classifiers are fused according to the proposed fusion method to form the fused classifier.

Evaluations on artificial low-dimensional data as well as experiments in a collaborative human-robot scenario with a 7-DoF robot arm show that the fused classifier outperforms the individual base classifiers according to increased accuracy and decreased uncertainty. In addition, it is robust to misclassifications of classifiers that are known to be unreliable.