Neural Network-Based External Impact Force Estimation for Compliant Mobile Robots

Most of the conventional approaches to mobile robot navigation avoid any kind of contact with the environment or with humans. However, most distance sensors have a limited field of view, so that collisions cannot be fully avoided in practical mobile robot applications. At the same time, direct physical contact also can be used as a means of intuitive communication between a robot and humans. We present a whole-body sensory concept based on a 6-DoF force-torque sensor to detect interaction forces between the robot body and humans. To distinguish between external contact forces and forces that result from the accelerations of the mobile platform, we employ a novel neural network-based filtering approach. The network fuses information from an inertial sensor, odometry data and the 6-DoF force-torque sensor. Our approach allows the robot to detect and react to physical contact during autonomous motion. Extensive experiments with our robot Canny demonstrate the effectiveness of our approach.