A parsimonious nonlinear manifold learning algorithm based on Diffusion Maps (DMs) is employed in this work to identify the low-dimensional manifold embedding the fluidic pinball dynamics. Two-dimensional direct numerical simulations of the incompressible Navier–Stokes equations are performed to compute the viscous wake flow behind the fluidic pinball by varying the Reynolds number Re. Five different flow regimes are considered, spanning from steady symmetric (Re < 18) to fully chaotic (Re > 115) conditions. In the first step of the Diffusion Maps embedding, the minimum set of DMs reduced coordinates (eigenvectors) necessary to represent the flow dynamics in all the regimes is found by projecting the high-dimensional simulation data into the reduced low-dimensional space (restriction operation). The nonlinear manifold lying in the state-space spanned by the three leading DMs coordinates is thus obtained by varying the Reynolds number Re, and its shape discussed in connection with the different physical mechanisms at play across all the flow regimes. Then, the time series embedded into the manifold are lifted back to the original space by means of Geometric Harmonics (lifting operation), such as to evaluate the reconstruction error between the “ground truth” solution (i.e. high-fidelity simulation data) and the DMs “reconstruction”. The performance of the DMs-based reconstruction is finally compared with a counterpart linear technique based on the Proper Orthogonal Decomposition (POD), which demonstrates the superiority of the proposed approach in parsimoniously representing the nonlinear dynamics up to chaotic conditions. The present work intends to be the first step towards the identification of a fully data-driven parsimonious DMs-based reduced order surrogate model of the fluidic pinball configuration suitable for its bifurcation analysis.