Shape Reconstruction of a Soft Actuator Based on Bezier Curves Using Soft Strain Sensors

Shape reconstruction of deformable actuators by means of intrinsic sensors is highly crucial for an accurate control of soft robots. Given the unstructured nature of the intrinsic sensors employed, a common approach is to leverage recurrent neural networks to estimate the position of a number of points along the main axis of the actuator. The shape is, then, reconstructed by fitting a kinematic model to the estimated points on the actuator. This paper proposes an alternative method in which a parameterized curve is chosen to model the deformation of the actuator. Feedback from the intrinsic sensors are utilized to directly estimate the model parameters by means of a neural network. The performance of the proposed approach was tested on a setup with an array of soft strain sensors attached to a tendon-driven actuator. The experiments were configured so that the actuator interacts with an external environment that applies a variable load on the deforming body, inducing a significant variation in the curvature of the backbone of the actuator. The proposed approach achieved an average estimation error of 1.16 mm in the tip position (0.6% of the actuator length) and 1.2 degrees in the tip orientation (less than 1% of the maximum tip orientation).