In this paper, we tackle the problem of multiconnectivity for latency critical industrial IoT services in 5G and Beyond. Sending multiple copies of the same packet on different radio interfaces is a popular mean for reducing delay. However, systematic redundancy may lead to congestion and more dynamic scheduling schemes such as Join the Shortest Queue (JSQ) are not guaranteed to be optimal. In order to derive the optimal multi-connectivity scheme, we first develop a novel discrete-time queuing model for the system, that integrates the heterogeneity of radio conditions and can consider any multi-connectivity policy. We then formulate an associated Markov Decision Process (MDP) that learns, online, the optimal state/action combination. Our simulations demonstrate that the derived optimal scheme outperforms classical ones like systematic redundancy and JSQ and is able to adapt dynamically to different traffic conditions. Interestingly, our proposed scheme yields a clear advantage when the objective is to ensure a hard latency constraint compared with the case where the average delay is to be minimized.