The paper deals with the specific control problems of the robot ALDURO, a four-legged walking machine, which is supposed to operate in unstructured terrains. The robot has four degrees of freedom at each leg, what makes impossible to the onboard operator directly to control the whole robot movement; therefore this task was partially automated. Hereby the collision avoidance task follows as a natural need, which must be implemented accordingly to ALDURO's reality: large dimensions, slow and spatial movements, unstructured environment, and no need of a long term path planning (there is an onboard operator). Considering such characteristics, it was decided to implement a reactive system, using a local map, based on the data from ultrasonic sensors. Such sensors are quite precise with respect to range measures, but suffer of intrinsically poor angular resolution, what conversely brings an advantage: they cover a whole volume at each measurement. Because of such angular inaccuracy, the inverse sensor model plays a relevant role to interpret each measure based on the sensor characteristics. The so formed information has to be added in an appropriated way to a base of knowledge (the local map), that is the so called data fusion process. Here is proposed a fusion by the use of a fuzzy inference system. Two kinds of linear functions were tested as output functions: constants and planes. The first one has one parameter per rule while the second, three parameters. That makes the simulations much slower when using planes, because the dimensions of the matrices employed at the calculation are directly proportional to the number of parameters per rule. To the adjustment of the parameters was used of Recursive Least Squares Method. Simulations were run with different number of partitions and different input terrain; as expected the system seems to be a little sensitive to different combinations of terrain and number of partitions, but the general performance holds.