Human neonates show a natural predisposition towards biological motion: despite the limited visual information available, they can distinguish the movement of other living agents from object motion. This ability has been suggested to be the basis for identifying conspecifics from birth, hence representing a fundamental skill for the development of social interaction. Inspired by this, we propose a computational model to detect biological motion in video sequences acquired by the iCub humanoid robot, as a first step to improve its action understanding and interaction capabilities in a developmental perspective. We propose a temporal multi-resolution motion description which automatically copes with different dynamics and builds on top of low-level features that capture biological motion regularities. The starting point of the representation is the optical flow, a low-level measurement which simulates the limited amount of visual information available at birth. The description is evaluated in its capability to discriminate between biological and non-biological movements. We show how a temporal multi-resolution descriptor can effectively deal with dynamic events of variable temporal duration and complexity. Then we provide evidence that our approach guarantees a robust classification of biological versus non-biological video sequences, which generalizes to new scenarios, including new agents and new actions, also in presence of severe occlusions.