The techniques adopted in "classical" AO (sensing in the visible, imaging in the near-IR) limit the achievable raw PSF contrast to about 10 5 in the central arcsecond. On nearby stars, this level is far from the theoretical PSF contrast limit imposed by photon noise in the wavefront sensor. A comparative study between wavefront sensing strategies shows that a focal-plane based wavefront sensor (WFS), combining wavefront sensing and scientific imaging on the same detector, seems optimal for high contrast imaging. This approach combines high WFS sensitivity, immunity to aliasing, chromaticity, and non common path errors and optical design simplicity. We show that such a system can be efficiently used as a second stage "Extreme-AO" system after a low-order AO system. The images acquired by the science camera are then used to drive the high-order DM (which also introduces the phase diversity needed for focal plane wavefront sensing). This scheme offers much flexibility, with the proper DM updates, the focal plane images can be simultaneously used to solve for the entrance wavefront and the presence of companions (which are incoherent with the speckles) below the speckle noise level. Control and data reduction algorithms are presented, as well as possible optical designs incorporating a coronagraph. A laboratory demonstration of this technique is currently being done at Subaru Telescope with a 1024 actuators MEMs DM. This experiment serves as a prototype to plan and design a similar system for Subaru's upcoming HiCIAO instrument (near-IR coronagraphic imager for adaptive optics).