This paper presents a complete framework (estimation, identification and control) for the implementation of joint-torque control on the humanoid robot HRP-2. While torque control has already been implemented on a few humanoid robots, this is one of the first im- plementations of torque control on a robot that was originally built to be position controlled. The challenge comes from both the hard- ware, which does not include joint-torque sensors, and the software interface, which only provides as command the desired joint angles (no motor current/voltage control). The contribution of the paper is to provide a complete methodology that is very likely to be reproduced as most robots are designed to be controlled only from their position. Additionally, the method is validated by an exhaustive experimental report including a comparison with the original position controller. The torque control is able to better track a reference trajectory while using lower values for the feedback gains (up to 25%). We also show that our controller can accurately track a desired Cartesian force.