Wireless localization is a crucial technology for determining the spatial position of nodes within a wireless network, with applications ranging from indoor navigation to asset tracking and robotic movement. However, traditional localization methods often rely on specialized hardware, while other localization methods suitable for commodity hardware tend to have limited accuracy. To improve localization accuracy, we propose models that integrate machine learning (ML) techniques for distance and position estimation. We observe improved localization performance by incorporating not only signal strength indicators from the physical layer, but also features specific to the mesh link layer of wireless mesh networks as model inputs. Furthermore, a proposed end-to-end machine learning localization model demonstrates promising performance by leveraging diverse network features for the direct inference of node position estimates. We evaluate the models using experimental data from a measurement campaign we conducted in an office environment hosting a wireless mesh testbed. Our method achieves an absolute position error of less than two meters. Using both information specific to mesh networks and novel machine learning methods yields improved performance, making our approach a promising solution for industrial applications.