The existing inverse optimal methods for canonical nonlinear systems assume that the system is modeled precisely and accurately, but dynamic uncertainties commonly exist and are unavoidable and difficult to model in practical engineering and industrial systems. This work removes this limitation and solves the problem of inverse optimal adaptive control for canonical nonlinear systems with dynamic uncertainties. Technically, a criterion on inverse optimality under dynamic uncertainties is newly proposed based on a new auxiliary system and a meaningful cost functional. With the new criterion, a robust adaptive fuzzy inverse optimal control scheme is proposed to design an inverse optimal controller, which, however, is not necessarily a stable controller. To solve this issue, a projection-based adaptation law is proposed to update the inverse optimal controller. Then, a small-gain approach is proposed to construct the links between inverse optimality and stability and to render that the closed-loop system is input-to-state practically stable. The proposed methods are successfully applied to industrial robots for demonstrations.