The mammalian spinal cord contains a locomotor central pattern generator (CPG) that can produce alternating rhythmic activity of flexor and extensor motoneurones in the absence of rhythmic input and proprioceptive feedback. During such fictive locomotor activity in decerebrate cats, spontaneous omissions of activity occur simultaneously in multiple agonist motoneurone pools for a number of cycles. During these 'deletions', antagonist motoneurone pools usually become tonically active but may also continue to be rhythmic. The rhythmic activity that re-emerges following a deletion is often not phase shifted. This suggests that some neuronal mechanism can maintain the locomotor period when motoneurone activity fails. To account for these observations, a simplified computational model of the spinal circuitry has been developed in which the locomotor CPG consists of two levels: a half-centre rhythm generator (RG) and a pattern formation (PF) network, with reciprocal inhibitory interactions between antagonist neural populations at each level. The model represents a network of interacting neural populations with single interneurones and motoneurones described in the Hodgkin-Huxley style. The model reproduces the range of locomotor periods and phase durations observed during real locomotion in adult cats and permits independent control of the level of motoneurone activity and of step cycle timing. By altering the excitability of neural populations within the PF network, the model can reproduce deletions in which motoneurone activity fails but the phase of locomotor oscillations is maintained. The model also suggests criteria for the functional identification of spinal interneurones involved in the mammalian locomotor pattern generation.