• machine learning
• object recognition
• multimodal sensing
• soft-gripper

The compliance and deformability of soft grippers allow adaptation to heterogeneous objects in shape and size. This work aims to achieve an active object picking, that exploits proprioception and contact sensing to perform object recognition. To this end, an LSTM learning model has been proposed as an alternative to state-of-the-art methods, achieving comparable performance both on a physical simulation and on a commercial artefact to recognize more than fifteen objects that differ in shape and size. We have also investigated the importance of sensory multimodality rather than a single source of information, how sensors’ spatial patterns and temporal dependencies influence the overall estimation, and the minimum training set of grasps per object to achieve good learning generalisation.