• active compliance
• control
• ESN
• LSTM
• machine learning
• soft robotics
• tendon-driven

Soft robots are extremely challenging in modeling and control due to their high dimensionality. The development of learning based controllers offers a wide range of possibilities. In this work a learning based controller for the active compliance of a tendon-driven soft arm has been developed using pyElastica as simulation environment. Recurrent neural networks, have been employed for the generation of the inverse kinematics (IK) of the soft manipulator and a comparison between Long-Short-Term-Memory (LSTM) and Echo-State-Network (ESN) has been performed. The control system, given the application of an external force at the tip of the robot, generates a trajectory to be followed to achieve active compliance. Comparing the movement of the robot induced by the same force with or without the controller, the results demonstrated a reduction of the force sensed at the tip of the robot, thus the soft manipulator has been able to follow actively the trajectory imposed by the external force.