Shapes reconstruction bridges real objects and their computer models. Most of the shape reconstruction techniques were derived for computer vision applications. A very important sense of human, tactile sensing can be applied to acquire shape information about 2D and 3D objects. Nevertheless, tactile data usually has a lot of noise. In this thesis, I present an applicable scheme that acquires shape data using a simple joystick sensor and then reconstructs 2D shapes and 3D patches. The 2D shapes are tracked by an Adept Cobra robot and represented as polynomial functions determined by the 3L fitting algorithm. The 3D shapes are composed of multiple patches, each of which is described by a polynomial function generated by least-square fitting. Experiments have been carried out with the robot. A display environment for 3D objects has also been developed.