Learning human-like facial expressions for Android Phillip K. Dick

Abstract

Android robots with human-like appearance can engage in superior conversational interaction with humans due to their ability to generate expressive faces. In this paper we describe a learning algorithm to map the rich facial features of a human to an Android, capable of replicating them. The methodology we employ is automated, marker-less, and has a high degree of generality. As a result, it can be applied to any Android face design, not just those exploiting linear or decoupled relationships between actuator inputs and facial feature point motions. To ensure robust results, we employed a genetic algorithm to search the actuator space and generate a rich set of robot facial expressions. Experimental results clearly show that once learning is complete, accurate facial expressions for the Phillip K. Dick Android can be achieved by mimicking those of a human.