Object-centric process discovery (OCPD) constitutes a paradigm shift in process mining. Instead of assuming a single case notion present in the event log, OCPD can handle events without a single case notion, but that are instead related to a collection of objects each having a certain type. The object types constitute multiple, interacting case notions. The output of OCPD is an object-centric Petri net, i.e. a Petri net with object-typed places, that represents the parallel execution of multiple execution flows corresponding to object types. Similar to classical process discovery, where we aim for behaviorally sound process models as a result, in OCPD, we aim for soundness of the resulting object-centric Petri nets. However, the existing OCPD approach can result in violations of soundness. As we will show, one violation arises for multiple interacting object types with loops that arise in collaborative systems. This paper proposes an extended OCPD approach and proves that it does not suffer from this violation of soundness of the resulting object-centric Petri nets. We also show how we prevent the OCPD approach from introducing spurious interactions in the discovered object-centric Petri net. The proposed framework is prototypically implemented.