Detecting Motion in the World with a Moving Quadruped Robot

For a robot in a dynamic environment, the ability to detect motion is crucial. Although legs are arguably the most versatile means of locomotion for a robot, and thus the best suited to an unknown or changing domain, existing methods for motion detection either require that the robot have wheels or that its walking be extremely slow and tightly constrained. This paper presents a method for detecting motion from a quadruped robot walking at its top speed. The method is based on a neural network that learns to predict optic flow at each timestep, thus allowing anomalies to be detected. The system is demonstrated to be capable of detecting motion in the robot’s surroundings.