An iPhone Pro is All You Need: Mimicking Facial Expressions on an Android Robot Head

Robots with very human-like faces are expected to be able to display human-like facial expressions. However, manually defining the animations of such facial expressions can be intricate and, thus, methods to automatically animate the actuators of android robot heads are promising. An approach to learn a mapping from human faces to android actuator configurations in a self-supervised way is presented here. Compared to previous approaches, convincing results are achieved with relatively few samples of train data. A capable technology to automatically extract facial expressions from vision input is identified as an important requirement for such approaches. The learned mapping is compared to a manual one in a user study and both are rated mostly positive. In effect, combining both is likely to result in even better mimicry capabilities.