Learning-Adaptive Deadband Sampling for Teleoperation-based Skill Transfer over the Tactile Internet

In remote skill transfer, demonstrations of a task are provided over a network via teleoperation and the remote robot learns from these teleoperated demonstrations. In a typical bilateral teleoperation scenario, transmission of position/velocity and force/torque samples require high packet rates for system transparency. In this paper we present a data rate efficient approach in teleoperation while ensuring robust remote learning from demonstrations. Our approach adapts the deadband parameter in the perceptual deadband-based kinesthetic data reduction method considering the confidence in the learned model. Our experimental results show that the mean packet rate to achieve the same quality of learning is drastically reduced when using the proposed approach.