Data-Driven Posture Control of Tensegrity Manipulator Based on Strut’s Inclination Angle

Abstract

Tensegrity structures have been utilised in soft robotics due to their flexible and lightweight nature. However, unlike traditional robots, these structures lack joint angles, which makes it challenging to use conventional angle sensors, and thus estimating the posture of the robot remains a challenge. To overcome this, we propose a data-driven approach for posture control of a redundant tensegrity manipulator using inclination angles of all struts, these angles are calculated relative to the gravitational direction. Specifically, we train a simple feedforward neural network (NN) to approximate a mapping from the inclination angles to the control inputs with a conditioning layer-wise averaged pressures. This network acts as a posture controller, mapping the desired inclination angles to the corresponding control inputs with conditioning by average pressure in the layer. The desired inclination angles corresponding to the desired posture can be obtained by demonstrating the robot in a direct teaching manner. We used the tensegrity manipulator consisting of 20 struts and 40 actuators to validate our approach. The experimental results showed that the tensegrity manipulator can reproduce the desired demonstrated postures.