Off-road mobile robotics arouses interest in many fields of application such as agriculture or surveillance where the need for several robots collaborating to cover an area is real. In this paper, the investigation carried out is the decentralized control of a formation of several wheeled mobile robots traveling along a path. The trajectory to follow is known, computed off-line or learned (beforehand or on-line from a manually-driven leader), and set as reference for the control of the formation. From an extended kinematic model of a robot, exact transformations permit to obtain a laterally and longitudinally decoupled model accounting for the bad grip conditions. Hence, the control of the steering angle and the velocity of the robot are derived separately, in order to ensure the desired formation shape. Indeed, the formation shape is defined by the lateral desired offset of each robot to the path and the longitudinal desired distances between a robot and its follower alongside the path (known as curvilinear distances). As the control of the lateral distance to the reference trajectory has been previously studied, the focus is particularly set in this paper on the control of the velocity of the robots to regulate the longitudinal distances between robots. It is based on an adaptive model to account for sliding by feeding the extended kinematic model, and a predictive control algorithm to anticipate for the velocity actuators delays. Experimental tests are driven to demonstrate the capabilities of the proposed approach.