With various industrial wireless networks greeting booming development, modern industrial devices configured with several network interfaces increasingly become the norm. Such multihomed industrial devices can increase application throughput by making use of multiple network paths, enabled by the multipath transmission control protocol (MTCP) (MPTCP). However, MPTCP might be challenged in the heterogeneous industrial networks because concurrent transmitting industrial application data over asymmetric network paths with different delays is almost bound to the receive buffer blocking problem, which is caused by out-of-order packet arrival and is harmful to the performance of the multipath transmission. The existing MPTCP solutions generally use static mathematical models to evaluate path quality and prohibit transmission on paths with poor quality, which are unable to perform efficiently under highly dynamic and complex network environments. Therefore, in this article, we propose a learning-driven latency-aware MPTCP variant, called ${l}\,^2$-MPTCP, which seeks to possibly mitigate the out-of-order packet arrival and receive buffer blocking problems associated with the network heterogeneity in the industrial Internet. ${l}\,^2$-MPTCP accurately computes each MPTCP path’s forward delay and assigns application data to multiple paths according to their calculated forward delay differences by using a novel multiexpert learning-enabled forward delay estimator. ${l}\,^2$-MPTCP dynamically manages path usage and chooses the optimal path collection for bandwidth aggregation and multipath transmission by using a promising reinforcement learning-empowered multipath manager. Experimental results demonstrate that ${l}\,^2$-MPTCP outperforms the current MPTCP solutions in terms of multipathing service quality.