An interactive inverse kinematics approach robustly generates natural poses in a large human-reachable space. It employs adaptive kd clustering to select a representative frame set from a large motion database and employs sparse approximation to accelerate training and posing. Model training is required only once. IK algorithms are fundamental in computer animation. However, designing energy functions that can generate natural poses for traditional IK algorithms is difficult. Style-based IK solves this problem by learning a prior model from motions. However, it might fail to generate natural poses when the desired poses differ considerably from the limited training poses. As we've shown, NAT-IK overcomes these limitations. It can relieve animators from time-consuming, back-and-forth, IK-pose adjustment. So, it's useful in automated applications such as games and virtual worlds