| dc.description.abstract | Robot kinematics modeling has been one of the main research issues in robotics research.The goal of this thesis is to endow a 3-D printed humanoid robot arm with the ability of positioning its fingers to a target position in real time. To achieve this goal, the robot system has to seek a high efficiency solution to inverse kinematics modeling. In this thesis, an SOM-like inverse kinematics modeling method for humanoid robot arms is proposed. The principal idea behind the proposed modeling method is to discretize the work space of the robot arm into a cubic lattice consisting of N_x×N_y×N_z sampling points. Each sampling point corresponds to a reciprocal zone and is assigned one grid node, storing five different data items: the index vector of the sampling point〖▁W_j〗^i, the coordinate vector of sampling point〖▁W_j〗^c, the position vector〖▁W_j〗^x, the output vector〖▁W_j〗^θ, and the Jacobian matrix〖▁W_j〗^J. All these five data terms can be quickly learned by the proposed modeling method from acollected training data set in an SOM-like manner. The training data set can be constructed by either the uniformly discretization scheme or the real-life data generation scheme. The computations of the joint angles corresponding to a target position in the work space involve the following two steps. First of all, we search the reciprocal zone which is closest to the target position. Secondly, the joint angles are approximated by a linear Jacobian expansion of the transformation＂Θ(＂ ▁x) via the position vector 〖▁W_j〗^x, the output vector 〖▁W_j〗^θ, and the Jacobian matrix 〖▁W_j〗^J within the reciprocal zone.
The performance of the proposed SOM-like inverse kinematics modeling method was tested on a 3-D printed robot arm with 5 degrees of freedom(DOF). The first experiment was conducted on a simulated robot arm environment. An averag eerror of 4.95mm and 4.91mm could be achieved over the 20,800 training data and 9,072 testing data, respectively. In addition, we can use a multi-step method to improve the performance to achieve 0.37mm errors in 0.03milliseconds. | en_US |