Human-Robot Ergonomic Handover via Deep Neural Network Based Adaptive Integral Sliding Mode Control

In this paper, we propose a strategy for performing human-robot ergonomic handover in the case of partial knowledge of the robot dynamics and in absence of assumptions about the shape or mass of the object passed. In particular, we propose a strategy for generating the reference for the manipulator and a Deep Neural Network based Integral Sliding Mode control with Adaptive discontinuous gain, in the paper referred to as DNN-AISM. The DNN weights are adapted on-line according to update laws derived directly from the theoretical analysis, without relying on previously collected data. The proposal is experimentally assessed relying on a Franka Emika Panda robot and on an Xsens MTw IMU sensor, producing highly satisfactory results.