Vehicle Routing Problems (VRPs) play an important role in industry for managing and optimizing operations for fleets of vehicles. Our project aims to incorporate new objectives that pertain to social and environmental sustainability together with classical productivity related ones. One key goal is for example to avoid contributing to congestion in the road network. To tackle these objectives, we use learning-based methods, since these have proven to be efficient, effective and broadly applicable. In previous work, we have explored new machine learning models to deal with the fundamental properties of the problems of interest, namely that they can be multi-objective with path flexibility. Through this project, we aim to continue our previous work by: 1) Further applying our two methods, GMS and NEPF, to increasingly complex and intricate settings. The goal is to demonstrate robustness, scalability and efficiency in realistic multigraph (multi-objective) routing setups. 2) Investigating non-autoregressive methods for multigraph multi-objective VRPs. Potentially, this family of methods could be more suitable for our specific setup. 3) Benchmarking against strong classical algorithms. We aim to demonstrate the advantages of our learning-based methods more clearly. Main supervisor: Balázs Kulcsár, Chalmers University of Technology, Department of Electrical Engineering.