Abstract

Multi-fingered hands can achieve dexterous manipulation by utilizing tactile sensing feedback. Abundant tactile information is useful specifically for recognizing contact states and plan motions with a synchronous motion of fingers. However, switching dexterous motions with high density tactile sensors is difficult and thus many studies have achieved one-single motion. In this paper, a neural network-based motion switching controller to achieve blind switching manipulation based on tactile feedback. An AE-LSTM model that combined restrictions on loss functions to improve the ability to switch movements and an attention mechanism that modulated attention to modalities required for movement was proposed. The success rate of manipulation for each model in unlearned initial positions and with unlearned objects was investigated. We confirmed that the proposed method exhibited the highest success rate of in-hand manipulation.