Deductive, mode-estimation has become an essential component of robotic space systems, like NASA's deep space probes. Future robots will serve as components of large robotic networks. Monitoring these networks will require modeling languages and estimators that handle the sophisticated behaviors of robotic components. This paper introduces RMPL, a rich modeling language that combines reactive programming constructs with probabilistic, constraint-based modeling, and that offers a simple semantics in terms of hidden Markov models (HMMs). To support efficient real-time deduction, we translate RMPL models into a compact encoding of HMMs called hierarchical constraint HMMs (HCHMMs). Finally, we use these models to track a system's most likely states by extending traditional HMM belief update.