The flexibilization and intelligent design of multi-variant production systems plays a major role in mass customization. One major difficulty in such manufacturing systems is often the rapid retooling of production lines and workstations in order to respond as flexibly as possible to changing customer requirements and variants. In most cases, this also includes reprogramming of machines and robots, which have to be adapted to new work sequences. However, such a reprogramming overhead can be effectively reduced by endowing the workstation with the ability to autonomously understand the dynamic changes on the assembly sequence. The goal of the present work is to present a concept to automate such a reconfiguring process based on machine learning models and artificial intelligence. The proposed paradigm aims to highlight the role of the human-robot cooperation on the effective achievement of flexible multi-variant manufacturing in shared collaborative workspaces. In particular, the underlying problems of scene understanding, tasks modelling and robot autonomy are deconstructed and discussed. In other words, the autonomous reconfiguration is defined in terms of the process awareness of the collaborative robot through the monitoring, identification and prediction of the environment and operator states in terms of previous, current and predicted assembly steps. Modelling the manufacturing process allows to abstract the spatiotemporal relations between the assembly steps of each variant’s sequence, independent of the operator actually executing the task. As a result, a flexible multi-variant assembly is obtained by the dynamic and intelligent adaptation of the collaborative robot integrated in the assembly workstation. The concept is partially implemented through an experimental setup of a collaborative assembly station in a learning factory lab.