12Bridging brain-scale circuit dynamics and organism-scale behavior is a central challenge in 13 neuroscience. It requires the concurrent development of minimal behavioral and neural circuit 14 models that can quantitatively capture basic sensorimotor operations. Here we focus on 15 light-seeking navigation in zebrafish larvae. Using a virtual reality assay, we first characterize how 16 motor and visual stimulation sequences govern the selection of discrete swim-bout events that 17 subserve the fish navigation in the presence of a distant light source. These mechanisms are 18 combined into a comprehensive Markov-chain model of navigation that quantitatively predict the 19 stationary distribution of the fish's body orientation under any given illumination profile. We then 20 map this behavioral description onto a neuronal model of the ARTR, a small neural circuit involved 21 in the orientation-selection of swim bouts. We demonstrate that this visually-biased 22 decision-making circuit can similarly capture the statistics of both spontaneous and contrast-driven 23 navigation. 24 25 30 be distinguished. If they signal an immediate threat or reward (e.g. the presence of a predator or 31 a prey), they may elicit a discrete behavioral switch as the animal engages in a specialized motor 32 program (e.g. escape or hunt, Budick and O'Malley (2000); Fiser et al. (2004); Bianco et al. (2011); 33 McClenahan et al. (2012); Bianco and Engert (2015)). However, most of the time, sensory cues 34 merely reflect changes in external factors as the animal navigates through a complex environment. 35 These weak motor-related cues interfere with the innate motor program to cumulatively promote 36 the exploration of regions that are more favorable for the animal (Tsodyks et al., 1999; Fiser et al., 37 2004). 38 A quantification of sensory-biased locomotion thus requires to first categorize the possible 39 movements, and then to evaluate the statistical rules that relate the selection of these different 40 1 of 24 Manuscript submitted to eLife actions to the sensory and motor history. Although the probabilistic nature of these rules generally 41 precludes a deterministic prediction of the animal's trajectory, they may still provide a quantification 42 of the probability distribution of presence within a given environment after a given exploration 43 time. 44 In physics terms, the animal can thus be described as a random walker, whose transition 45 probabilities are a function of the sensory inputs. This statistical approach was originally introduced 46 to analyze bacteria chemotaxis (Lovely and Dahlquist, 1975). Motile bacteria navigate by alternating 47 straight swimming and turning phases, so-called runs and tumbles, resulting in trajectories akin to 48 random walks (Berg and Brown, 1972). Chemotaxis originates from a chemical-driven modulation 49 of the transition probability from run to tumble: the transition rate is governed by the time-history 50 of chemical sensing. How this dependency is op...