Existing visually guided walking machines have difficulty traversing terrain cluttered with obstacles. These walking machines use computationally intense approaches that require construction of a geometrically correct model of both the environment and the robot. However, most terrestrial vertebrates accomplish this task easily, suggesting that better strategies exist. We present a model inspired by recent research in cats and humans. In our model, perception and action are tightly coupled. The mapping is adaptive and based on experience. The goal of the adaptation is to use distance measurements to smoothly modulate a central pattern generator (CPG) controlling gait. A key element in our model is the use of a temporal gating hypothesis. This hypothesis simplifies the learning problem and is consistent with biological observations. Our approach does not require that a geometric representation of the environment be created or updated based on new observations. This is in strong contrast to current practice in machine vision and robotics of surface reconstruction as a prerequisite to planning. Our simulation results indicate that the desired mapping can be learned quickly with few mistakes before perfect performance is achieved. The resulting gait modulation is smooth and coordinated with the phase of the CPG controlling the robot.