This paper describes models and algorithms for intelligent control of a group of autonomous mobile robots, which perform object transportation task in complex environments. The proposed models allow avoiding obstacles in the way of the multi-robot system as well as motion coordination for several robots carrying a large-sized object. We use Q-learning to provide robots adaptability to unknown environments. The 2d-map data collected during system operation joined with actions for several robots serves as an input for the learning subsystem. The primary output is a value of estimated efficiency for the given control decision. The use of a convolutional neural network as a core element of the reinforcement learning module allows for efficient encoding of system operation situations, thus providing for a reliable and efficient control algorithm