Steering of Flexible Needles Using an LSTM Encoder with Model Predictive Control

This work presents a method of using recurrent neural networks (RNNs) in combination with model predictive control to determine an optimal control strategy for needle manipulation for deep tissue applications. The paper discusses creating a needle insertion model from experimental data that can then be used to generate data for training the proposed network. The RNN makes no assumptions about needle-tissue interaction, instead it learns the dynamics of the interaction from simulated and experimental data. It is shown in the paper how deep recurrent neural networks can create a simple cost function enabling model predictive control to determine an optimal sequence of needle manipulations. Simulation results show that the proposed control structure can accurately predict the effect of current control actions on future trajectory. Simulation results indicate that the proposed control strategy is able to determine an optimal control strategy within a few time steps of the simulation initializing, while requiring only one rotation to enable a needle to be steered to within 1.1mm of the desired target.