A non-linear Model for Predicting Tip Position of a PliableudRobot Arm Segment Using Bending Sensor Data

摘要

Using pliable materials for the construction of robot bodies presents new and interesting challenges forudthe robotics community. Within the EU project entitled STIFFness controllable Flexible & Learnable manipulatorudfor surgical Operations (STIFF-FLOP), a bendable, segmented robot arm has been developed. The exterior of theudarm is composed of a soft material (silicone), encasing an internal structure that contains air-chamber actuatorsudand a variety of sensors for monitoring applied force, position and shape of the arm as it bends. Due to the physicaludcharacteristics of the arm, a proper model of robot kinematics and dynamics is difficult to infer from the sensoruddata. Here we propose a non-linear approach to predicting the robot arm posture, by training a feed-forward neuraludnetwork with a structured series of pressures values applied to the arm's actuators. The model is developed acrossuda set of seven different experiments. Because the STIFF-FLOP arm is intended for use in surgical procedures,udtraditional methods for position estimation (based on visual information or electromagnetic tracking) will not beudpossible to implement. Thus the ability to estimate pose based on data from a custom fiber-optic bendingudsensor and accompanying model is a valuable contribution. Results are presented which demonstrate the utility ofudour non-linear modelling approach across a range of data collection procedures.