Sensorimotor systems are inherently uncertain, either because of the physical nature of the sensory stimulus or because of noise within neural circuits. To curtail such limitations, sensorimotor systems combine information about the same event across different sensory sources to create robust internal representations for behavioral performance. Here we show that congruent multimodal self-movement signals are consistently found throughout populations of visual neurons, the lobula plate tangential cells (LPTCs). LPTCs monitor head and/or body rotations to support steering during locomotion1-3. We investigated the computational purpose of this congruent interaction by contrasting activity in LPTCs in walking flies with the predictions from Bayesian integration of independent inputs4. Under varying sensory reliability and in the presence of perturbations, we found that multimodal signals within LPTCs are compatible with the predicted weighting of unimodal signals based on their reliability. The precision of self-motion estimation by LPTCs faithfully follows the predicted precision from optimal integration of the unimodal inputs. Furthermore, we found a signature of normalization within the activity of these cells, a canonical computation generally considered to increase the regime for optimal operation of neurons. To examine the role of this optimal integration on behavior, we developed artificial agents following the behavior of an exploratory walking fly. Exploratory flies change their ability to walk straight as a function of the visual reliability of the environment. This relationship is mimicked by artificial agents employing the measured multimodal integration rules even under perturbation conditions. Altogether, our findings demonstrate that the LPTC network optimally combines multimodal information to infer self-motion robustly and support steering under natural uncertainty. Our results further open the possibility to mechanistically study the neural implementation of general computations such as Bayesian inference and normalization during a continuous and dynamically changing behavior in a compact and genetically tractable system.