The mammalian striatum is critically involved in perceptual decision-making, where disrupted dopamine transmission leads to perceptual disturbances such as hallucinations. The precise mechanism by which striatal dopamine contributes to perceptual decisions is not known. We investigated this by developing a circuit model in which perceptual learning and value learning are implemented in parallel corticostriatal loops to inform action selection. Previous models of corticostriatal plasticity have focused primarily on how reward prediction errors signaled by dopamine enable value learning in the ventral striatum. Here, we used meta-learning to determine the dynamics of error signals that optimize biologically plausible learning of prior expectations about latent world states in the sensory striatum. We found that learning was optimized by sensory prediction errors, i.e., momentary fluctuations in the subjective belief about stimulus identity. When performing a signal detection task, this model adapts to stimulus-history and reward-history by changing its subjective beliefs and subjective policy respectively. We tested this in human participants and found a pattern of double dissociation in their behavioral responses identical to the model. At the neural level, the model predicts that stimulus-history should influence dopamine signaling in the sensory striatum but not the ventral striatum. We tested this prediction using data from mice performing auditory signal detection while having dopamine signals recorded from the sensory and ventral striatum and found signal-history effects only in the former. Furthermore, in alignment with the model, we found that stimulus-induced dopamine responses were modulated more strongly by stimulus identity and subsequent performance feedback in the sensory striatum and ventral striatum respectively. We propose that corticostriatal loops leverage dopamine heterogeneity to learn different types of environmental statistics along the dorsoventral axis to guide decision-making.