The Internet of Things and its applications can be seen in all aspects of human life, and adding mobility features to objects has brought about new challenges. Mobility causes dynamics in topology and instability in the network links, and introduces several hurdles in discovering the path with less overhead and low delay. Most of the routing algorithms are based on the on-demand distance vector routing algorithm based on the source node's request. In recent years, several improvements have been presented on this mechanism, and its multi-path version is one of the notable improvements. Due to the high dynamics of the network and the momentary changes in the mobile Internet of Things, it is impossible to predict all the circumstances and adjust the optimal parameters in advance. Thus, a novel routing strategy based on machine learning and multi-path distance vector routing is introduced in this research. Its goal is to assess network conditions and choose the optimal interface node for packet forwarding. Five machine learning algorithms, including decision trees, random forests, and support vector machines, are used in the proposed approach for learning and evaluation. A number of network parameters, including node velocity, the number of neighboring nodes, buffer size, remaining energy, and the average distance between each node and its neighbors, are taken into account when choosing the best interface node. Distance of each node with its neighbors are considered for selection of the appropriate interface node. The simulation results conducted through Python indicate that the decision tree and the gradient boosting have the best results in the collected data set, and by combining them with the proposed approach, the end-to-end delay is reduced by 30%, and the number of lost packets is reduced by 18%.