Intention Recognition with Recurrent Neural Networks for Dynamic Human-Robot Collaboration

Abstract

A new method to recognize the intention of a human worker while performing a collaborative task with a robot is proposed. For this purpose, two recurrent neural network (RNN) architectures capable of predicting the worker's target were developed. The first uses marker-based tracking of hand positions and the second RGB-D videos of human motion. The system was implemented to perform a collaborative assembly task. The results show high intention prediction accuracy for both networks, with accuracy increasing once a larger portion of human motion has been observed, making the proposed method viable for efficient and dynamic human-robot collaboration. Furthermore, we developed a framework that enables online adaptation of robot trajectories based on estimated human intentions.