Beyond the fact that players don’t take at the first node, the experimental literature on the centipede game has highlighted a number of puzzles that neither the standard rational approach or existing behavioural theories can perfectly explain. In this paper, I propose a model of heterogeneous levels of altruism and ambiguity to shed light on these puzzles and explain the experimental data. This model yields a unique equilibrium in which players use pure strategies. Furthermore, a uniform distribution of altruism and ambiguity fits better the centipede game behaviours than the two-player type model of McKelvey and Palfrey (1992). To best fit the data, around 90% of the players should be modelled as ambiguity-averse and only 10% as altruistic.