Variable Duration Movement Encoding with Minimal Intervention Control

Programming by Demonstration (PbD) offers a user-friendly way of skill transfer from human to robot. Typically, demonstration data do not contain the control inputs required to reproduce the demonstrated skill. These control commands can be obtained from a low-level controller tracking the movement dynamics of the skill. Both the generation of the reference signal and the selection of the low-level controller are critical for reproduction. In this work we present a low-level controller with a minimal intervention strategy based on Model Predictive Control (MPC). MPC is an anticipative control strategy that allows to handle perturbations online. The control objective required for MPC is obtained from a movement model encoding the demonstrated skill. In contrast to our previous work, where the generation of the control objective explicitly depends on a time signal we propose a method that does not have this explicit time dependency. Instead, we propose to model temporal behavior of the movement as a probability distribution using a Hidden Semi-Markov Model (HSMM). This creates a model that is able to encode and reproduce movements with variable temporal duration. We compare the method in simulation with a time-driven motion encoding. In addition, we demonstrate the proposed method in a real-world robot experiment.