Upright balance control is a fundamental skill of bipedal robots for various tasks that are usually performed by human beings. Conventional robotic control is often realized by developing accurate dynamic models using a series of fixed torque-ankle states, but their success is subject to accurate physical and kinematic models. This can be particularly challenging when external disturbing forces present, but this is common in unstructured robotic working environments, leading to ineffective robotic control. To address such limitation, this paper presents an adaptive ankle impedance control method with the support of the advances of adaptive fuzzy inference systems, by which the desired ankle torques are generated in real time to adaptively meet the dynamic control requirement. In particular, the control method is initialised with specific external disturbing forces first representing a general situation, which then evolves whilst performing in a real-world working environment by acting on the feedback from the control system. This is implemented by initialising a rule base for a typical situation, and then allowing the rule base to evolve to specific robotic working environments. This closed loop feedback and action mechanism timely and effectively configures the control system to meet the dynamic control requirements. The proposed control method was applied to a bipedal robot on a moving vehicle for system validation and evaluation, with robotic loads ranging from 0 to 1.65 kg and external disturbances in terms of vehicle acceleration ranging from 0.5 to 1.5 m/s, leading to robotic swing angles up to 7.6º and anti-disturbance timespans up to 8.5 s. These experimental results demonstrate the power of the proposed upright balance control method in improving the robustness, and thus applicability, of bipedal robots.