We aim at addressing in this paper the issue of non-verbal communication involved in a human-robot interaction. We first propose a taxonomy of assistance that an operator may expect from a partner during a long object manipulation task. Then, this paper presents a method to detect which kind of robot assistance is needed, thanks to the decoding of haptic cues conveyed by the operator. An experimental study on thirty subjects moving a long object together with a robot that is providing an assistance is introduced. A statistical analysis highlights the relationships between haptic measures and intentions of motion. This analysis shows that wrench measurements constitute an incomplete information to detect the operator's intention of motion.