To work with humans, robots must discriminate between movements of living and non-living things — a challenging task. Alessia Vignolo at the Italian Institute of Technology in Genoa and her team developed a machine-learning algorithm that exploits the characteristic smoothness with which humans change speed and trajectory on a split-second timescale.
By observing how sharply motion changed, the algorithm learned to differentiate between human actions, such as rolling dough, and the movements of inanimate objects, such as toy trains. The program does not interpret objects on the basis of their appearance, so it was successful even when viewing unfamiliar actions and when a scene was partly obscured.
Installed in a humanoid robot, the algorithm was able to automatically direct the machine’s gaze at humans, a useful skill for social interaction, say the authors.