In the unstructured task environment, the collaborative reconstruction of multiple mobile robots to form combined robots is significant to improve the task adaptability. However, in general, the progress in modeling and control for enabling the practical deployment of combined robots is still insufficient. Based on nonholonomic constraints and the Lagrange method, the kinematics and dynamics of combined robots at arbitrary reconstruction scales are established. At the kinematic level, a smooth obstacle avoidance strategy based on adaptive boundary compensation safety geofences is designed to ensure the safety during movement. At the dynamic level, the adaptive cooperative fault-tolerant mechanism is introduced to ensure the rationality of the torque distribution of the robot system and avoid performance degradation, to enhance the fault-tolerant ability during motion. At the same time, a new filter extended state observer (FESO) is proposed, which can realize robust tracking control of velocity under unknown lumped disturbance by suppressing high frequency oscillation caused by noise. Finally, a real combined robot experiment platform is built to further verify the fault tolerance, robustness and safety of the proposed method.