Learning robotic needle steering from inverse finite element simulations

Tissue motion compensation during robotic needle steering is a challenging research topic. While the deformable non-linear coupling between needle and tissue is captured by simulation-based control strategies, they increase significantly the computational cost of the control. In this work, we rely on machine learning methods to enable autonomous robotic needle steering with very low computation times. We propose to use an Extreme Learning Machine (ELM) to learn an inverse model which accounts for needle-tissue interaction. The ELM trains with synthetic data generated from multiple needle insertions controlled by inverse finite-element simulations. Results indicate the method is able to achieve clinical compatible precision, and it's robust to previously unseen trajectory-shapes and variable tissue elasticity parameters, while using only a third of the computational time demanded for simulation-based methods.