The recent advances in Integrated Sensing and Communications (ICAS) essentially transform mobile radio networks into a diverse, dynamic and heterogeneous sensing network. For the application of localization, the acquired sensing data needs to be processed to estimates of the state vectors of the targets in a timely manner. This paper aims at providing a roadmap for the development of a suitable computing system for ICAS. We propose to embed the signal processing into the concept of edge computing. It provides the necessary theoretical computing framework, since it alleviates the need for communication with a remote cloud. To obtain localization information in such a distributed, asynchronous and heterogeneous scenario, we study how existing maximum likelihood estimation techniques can be transformed into algorithms that can be orchestrated close to the edge. The advantage of these approaches is that they have well studied statistical properties and efficient algorithmic implementations exist. We propose to study how to derive a graph that encodes these algorithms’ processing by relating individual and isolated computations in terms of the input/output-behavior of so-called compute nodes. This compute graph structure can then be flexibly distributed across multiple devices and even whole processing/sensing units. Moreover, modern computing architectures leverage such graph structures to optimize the efficient use of computing hardware. Additionally, once this graph is constructed we have laid the groundwork for the possibility to exchange certain compute steps by deep learning architectures. For instance, this allows to sidestep some costly iterative part of traditional maximum likelihood estimators, which further contributes to the low-latency of the localization task. Moreover, deep learning methods bear the promise of being more robust to model mismatches in contrast to the conventional model based approaches. As a consequence, we can then study the relation between those classical methods and the new deep learning based methods and analyze the achievable performance. One key final result will be a deeper understanding of how well the maximum likelihood approach can be applied to ICAS and how much it profits from the combination with modern deep learning techniques.