A system, comprising a computer that includes a processor and a memory, the memory storing instructions executable by the processor to input second vehicle sensor data received via one or more of vehicle-to-vehicle (V-to-V) and vehicle-to-infrastructure (V-to-I) networking into a generative query neural network (GQN) trained to generate second vehicle viewpoint data and predict a path for a second vehicle based on the second vehicle viewpoint data with a reinforcement learning deep neural network (RLDNN). The computer can be further programmed to obtain a confidence for the predicted second vehicle path from a Bayesian framework applied to the RLDNN and determine a first vehicle path based on the predicted second vehicle path and the confidence.