For humanoid robots, reaching with their hands as much workspace as possible is an important issue, since the locations of the target objects are ranging from the floor to the place above robot's head. Furthermore, it is necessary to solve inverse kinematics for the whole body in real time to adapt to the constantly-changing environment. First, we propose a dual-arm inverse kinematics with a visual feedback mechanism. The proposed solution uses a weighted pseudo inverse matrix and a pipeline calculation. The weights improve the robot's balance without significantly increasing the calculation time. Second, we propose a method to achieve real-time motion generation for a humanoid robot by separating the inverse kinematics calculation into simpler problems. Using regression to estimate the torso orientation, we solve the inverse kinematics for the lower body and both arms independently. Based on the target pose of both hands as input, we calculate the orientation of the torso and determine the target position of the center of mass considering the reachability of both arms. At each control step, we calculate the joint angles of the lower body from the position of the center of mass, feet poses, and torso orientation. Finally, we calculate the joint angles of both arms. We experimentally verify the proposed method by generating motions to approach objects. We verify the effectiveness of the proposed method to keep the robot's balance using the ZMP. We also apply the proposed method to a human-size humanoid robot for reaching low-height positions while hunkering down. From the experimental results, we prove the following advantages of the proposed method. The proposed inverse kinematics solver is ten times faster than the numerical solution using the Jacobian matrix. The accuracy of the hand pose control and the calculation time are better than the conventional method. The error and the calculation time are lower than with the conventional method. The error using the proposed method is smaller when the target positions are at the far side in the lateral direction. We also verify the applicability of the proposed method by following a sequence of random hands' target positions.