Vertebrate locomotion is produced by spinal circuits, sensory feedback and body mechanics acting as one closed loop. Which of these components generates the rhythm, and in what proportion, has been debated for a century, and the question is hard to settle experimentally: removing a sensory pathway shows that it is necessary for a behavior, but not that any particular circuit is sufficient to produce it, because every other sensorimotor loop is still in place. This thesis addresses the problem computationally, using spiking neural networks coupled to simulated bodies, fluids and physical robots. The spiking level was chosen because its manipulations — silencing a defined interneuron class, removing a mechanotransduction channel, cutting synaptic connections — have direct experimental counterparts, so that models constrained by data can return falsifiable predictions.

Four studies are presented. The first introduces a spiking neuromechanical model of the salamander that reproduces fictive locomotion at the level of the full cord, hemicord, isolated segments and hemisegments, and produces both swimming and walking from a single architecture under modular descending drive. Intermediate levels of axial stretch feedback increase speed, lower the cost of transport, reduce intersegmental phase lag and enlarge the range of conditions under which the network functions, whereas excessive feedback makes it uncontrollable by descending commands. The second study takes the same model into a completely transected spinal cord with incomplete axonal regrowth. Open-loop networks recover rhythmicity but never the healthy locomotor phenotype; intermediate feedback below the lesion makes recovery faster and complete, on a time scale compatible with the experimental recovery of swimming, while excessive feedback restores movement immediately but in a pathological regime: restoring movement is therefore not the same as restoring healthy locomotion. The third study identifies a low-order spinal circuit for collective behavior in zebrafish, in which intraspinal proprioceptors deliver curvature-based glycinergic inhibition onto rhythm-generating V2a interneurons. A model constrained entirely by experimental data and a bio-inspired robot running the same controller in a real turbulent wake spontaneously adjust their tail beat phase according to their distance to neighboring fish, saving energy in the process. The computational results establish sufficiency alongside the necessity demonstrated with \emph{in vivo} experiments. The fourth study shows that adult zebrafish swim in two mechanically optimal modes, and that the longitudinal reach of two commissural inhibitory populations, rather than their targets, is the variable that selects between them.

Together these results support a specific view of locomotor control: central pattern generation, sensory feedback and body mechanics are partially redundant, the best-performing and most robust configuration is an intermediate one, and the spinal cord shapes the geometry of movement as well as its timing.