Controlling Soft Robotic Arms Using Continual Learning

Learning-based modeling and control of soft robots is advantageous due to neural network's ability to capture complex dynamical effects with low computational cost. Continual Learning techniques add further value to these methods by allowing networks to learn from continuously available data without incurring into catastrophic forgetting. In the context of soft robotic control, such capability can be exploited to design controllers able to continuously adapt to changes in robot dynamics, frequently due to material degradation or external interactions. This should be done without forgetting the control under normal working conditions which can be recovered as soon as the external interactions return to normal. In this paper elastic weight consolidation is used to continuously re-tune a neural network-based controller while changing the external loading of a soft robot. We demonstrate experimentally on a soft robot arm that this method outperforms plain stochastic gradient descent in tracking tasks, in the context of a continuously changing loading condition. We also show that the proposed control architecture can improve its performances when exposed to loading conditions already experienced. This paper represents a first step towards the introduction of continual learning methods in the soft robot control field