Uncertainty-Aware Haptic Shared Control With Humanoid Robots for Flexible Object Manipulation

We propose a haptic shared control system that predicts human manipulation intentions using a neural network and adaptively presents haptic guidance to achieve smooth robot control remotely. Although the haptic shared control has garnered increasing attention as a method to improve operability in remote operations, incorrect guidance can worsen operability. In this study, we dynamically switch the strength of haptic guidance presentation depending on the uncertainty of the inference results of the neural network. Thus, we weaken the haptic guidance presentation strength for predictions in which the neural network lacks confidence and strengthen it for those with high confidence, thereby achieving guidance presentation that does not impede human manipulation. As a result of experiments using the Nextage OPEN upper-body humanoid robot, in a task involving folding a flexible object, we succeeded in reducing task execution time by 17.1% compared to that with an existing method that determines the strength of haptic guidance presentation without considering the confidence of the neural network.