Continuum manipulator offers superior flexibility, deformability and adaptability to the environment, which makes it ideally suitable for safe interaction and for applications in confined spaces. Owning these advantages, the continuum manipulator gains increasing interests in the fields of surgical, underwater, inspection, etc. This paper introduces a novel continuum manipulator, which is designed as an independent and auxiliary modular for bringing extra dexterity and reachability to different rigid platforms. With the aim at precisely calculating the gripper pose of the manipulator, a probabilistic model-based approach is used, which learns a mapping among the actuator space and task space from experiments by using dynamic mixture of Gaussians. The learned model and the control approach are validated with the help of a 3D trajectory tracking system. Finally, to test the versatility and reliability of the manipulator, we mount the actuation base of the continuum manipulator to a rigid manipulator, which are then used to manipulate an object in a confined space. The results of the experiments show that the proposed continuum manipulator can be controlled effectively by using a learned probabilistic model and the dexterity and workspace of a robotic system could be enhanced significantly by soft-rigid arm collaboration.