Giovanni Franzese

In Learning from Demonstrations, uncertainties can lead to bad generalization of the learned policy. My project proposes to interactively teach a robot from human corrections. The interaction between the human and the robot is made easier thanks to the implementation of variable robot compliance as a function of uncertain/ambiguous situations, in which more than one action have similar probabilities and avoiding a random action selection. The human feedback is used to correct the current policy database or acquire more data. The aim is to improve the user experience, the learning performance and safety. Different applications have been studied from learning variable impedance policies for assembling/disassembling, interactive task feature selection and complex picking policies at non-zero velocity. The proposed approaches exploited Gaussian Processes for policy learning, identifying regions of uncertainty as well as allowing interactive corrections, compliance modulation and active disturbance rejection.