Direct detection of planets of neighboring stars by imaging from the ground is extremely challenging, and requires bringing together and extending all that has been learned in the past two decades in adaptive optics experiments and development. From a distance of 30 light years, the planet Jupiter would be 10 9 times dimmer than the sun, and at 0.5 arcsec separation would be lost in the strong glare of scattered light from the central star. In this paper, we lay out the requirements for adaptive optics to allow direct detection with a large telescope. The stellar halo must be suppressed by several orders of magnitude, and speckle noise caused by correlated wavefront errors must be severely reduced to allow efficient smoothing of the halo through averaging of random fluctuations caused by photon noise. For a 6.5 m telescope imaging near 1 micrometer wavelength, suppression of the stellar halo to 10 -6 of the peak intensity allows direct detection of Jupiter-like planets in several hours of integration. A deformable mirror with approximately 10,000 correction elements is needed, updated at 0.5 millisec intervals using a wavefront sensor optimized for use with bright stellar sources. Local filtering of wavefront sensor data is required to overcome correlated errors arising from time delay between sensing and imaging. Correction of the strongest amplitude errors caused by scintillation allows the required integration times to be decreased by a factor of 2. We present results of detailed simulations for an adaptive system which achieves the above goals, for imaging at 1 micrometer wavelength with a 6.5 meter telescope. A simulated image of a solar system twin at 8 parsecs shows Jupiter at the 5 σ level for a 5 hour integration. We plan to develop and use a similar system to conduct a two- hemisphere survey of bright nearby stars on the twin 6.5 m MMT and Magellan telescopes.