Big vehicular data is anticipated to become the new fuel for catalyzing the further development of connected and autonomous driving. Vehicles themselves will act as mobile sensor nodes that actively sense their environment and gather meaningful data for novel crowdsensing-enabled services such as the distributed generation of high-definition maps, traffic monitoring, and predictive maintenance. However, the implied tremendous increase in massive Machine-Type Communication (mMTC) represents an enormous challenge for the coexistence of different resource-consuming applications and entities within the limited radio spectrum. A promising approach for achieving relief through a more resource-efficient usage of existing network resources is the utilization of client-based intelligence. Novel communications paradigms such as anticipatory mobile networking aim to improve decision processes within wireless communication systems by explicitly taking context information into account. In the context of vehicular crowdsensing, these methods exploit the delay-tolerant nature of the targeted applications for scheduling the data transfer with respect to the expected resource efficiency. If the current radio channel and network load conditions do not allow a resource-efficient transmission, the data transfer process is postponed and the acquired data is aggregated locally in favor of a better transmission opportunity in the near future along the expected vehicular trajectory. In the following, the different evolution phases of the novel Channel-aware Transmission (CAT) scheme are presented. These are characterized by a sequential introduction of different machine learning methods. While the basis CAT approach applies a probabilistic channel-access mechanism based on measurements of the Signal-to-Noiseplus- Interference Ratio (SINR), Machine Learning CAT (ML-CAT) applies supervised learning for predicting the currently achievable data rate using features from the network context, the mobility context, and the application context domain. This approach is then further extended by Reinforcement Learning CAT (RL-CAT) through the autonomous detection and exploitation of favorable transmission opportunities. Finally, Blackspot-Aware Contextual Bandit (BC-CB) integrates a priori knowledge about the geospatially-dependent uncertainties of the prediction model, which is uncovered by unsupervised machine learning. It is shown that machine learning-aided opportunistic data transfer is not only able to increase the average data rate of the individual transmissions; it also achieves a massive reduction of the occupied network resources and the power consumption of the mobile device. The price to pay is an increase of the Age of Information (AoI) of the sensor measurements. In addition to the presentation of the novel opportunistic data-transfer approaches, new machine learning enabled methods for simulating these anticipatory mobile networks are presented, discussed, and validated.