Data-oriented Inverse Kinematics Applied to Soft Robots With Fiducial Markers for Shape Feedback

Soft manipulators offer many advantages when operating in environments with humans and difficult-to-reach areas. Their high dimensionality introduces modeling difficulties to get an inverse kinematic model. This paper investigates data-oriented inverse kinematics models in a soft robot prototype using fiducial markers to get shape information. A missing values imputation flow is proposed to deal with non-detected ArUco Markers. The results demonstrate the potential of visual feedback combined with machine learning to assist the manipulator estimate control actions to reach a desired shape. Three neural network topologies are evaluated, among which are LSTM, MLP, and Transformer, with the last one returning the best performance.