AbstractRobots destined to tasks like teaching or caregiving have to build a long-lasting social rapport with their human partners. This requires, from the robot side, to be capable of assessing whether the partner is trustworthy. To this aim a robot should be able to assess whether someone is lying or not, while preserving the pleasantness of the social interaction. We present an approach to promptly detect lies based on the pupil dilation, as intrinsic marker of the lie-associated cognitive load that can be applied in an ecological human–robot interaction, autonomously led by a robot. We demonstrated the validity of the approach with an experiment, in which the iCub humanoid robot engages the human partner by playing the role of a magician in a card game and detects in real-time the partner deceptive behavior. On top of that, we show how the robot can leverage on the gained knowledge about the deceptive behavior of each human partner, to better detect subsequent lies of that individual. Also, we explore whether machine learning models could improve lie detection performances for both known individuals (within-participants) over multiple interaction with the same partner, and with novel partners (between-participant). The proposed setup, interaction and models enable iCub to understand when its partners are lying, which is a fundamental skill for evaluating their trustworthiness and hence improving social human–robot interaction.