Learning Contact Forces in Human-Wearable-Robot Interaction Using Morphological Computation

This paper presents a morphological learning framework to predict/estimate the contact forces between a human finger and a pneumatically actuated soft robotic digit by harvesting the reservoir computing power of the robot's soft bodies. During the actuation, the soft robot and human finger physically interact at multiple contact points; to predict the contact forces, a physical reservoir framework was established. The soft actuator segment was considered as the physical reservoir integrated into a linear readout. The deformation of the soft-bodied was visually tracked using an image processing code during actuation based on distributed attached markers while the force sensors measured the contact forces. The linear readout adds up the relative displacements to estimate the interaction forces. The readout weights were trained using a linear regression based on the desired output forces. On the other hand, a theoretical lumped model of the soft actuator segment, analogous to the markers, is developed using a network of mass-spring-damper. The model was simulated with experimentally measured input pressure force acting on the mass points while the spring length was calculated and added up using the linear readout. The physical reservoirs were trained for three individual soft actuators and exploited for estimating the contact forces. The effectiveness of the learning framework was experimentally and theoretically demonstrated, where it was able to estimate the actual contact forces between the soft robotic digit and the anthropomorphic finger model within an acceptable range of errors.