The conventional sense-and-avoid collision avoidance mode of UAV (unmaned aerial vehicle) lacks applicability and timeliness in a multi-threat environment. In this paper, a new efficient collision avoidance approach for uncertain threat environments derived from the idea of autonomous mental development is proposed. The proposed collision avoidance pattern consists of a sensory layer, a logic layer and a development layer. The threat information is sensed using the sensory layer, and the path planning approach in the logical layer is applied to the output configuration of UAV. In the development phase, the developmental networks approach is used for online learning, training and updating the logical layer so as to form the sense-action mapping, which is stored as the "basic experience" for UAV executing the avoidance manoeuvre. In the implementation phase, the command is executed by matching the sensing information and action base. The simulation results show that the proposed approach has better timeliness compared to the conventional approaches.