The deployment of intelligent reflecting surfaces (IRSs) to assist multiple-input multiple-output (MIMO) communications stands as a promising step in overcoming some of the limitations of current vehicular communication systems. In this paper, we focus on the joint optimization of the IRS and the precoders in the uplink of an IRS-assisted MIMO communication established between several connected vehicles and a roadside unit (RSU). We consider a realistic imperfect channel state information (ICSI) model and propose several adjustments to the deep contextual bandit-oriented deep deterministic policy gradient (DCB-DDPG) framework to address this complex optimization problem. Simulation results show that the proposed solution is a robust alternative because it keeps learning from interactions, even when the estimation error statistics are unknown.