Battery life. If necessary, humans can stay active for a long time. Robots can't. If it dies, you either manually get the robot, or leave it for the enemy. The more processing and sensors on a platform, the sooner it dies. An almighty AI will likely require an almighty amount of power.
Cameras. Humans swiftly adapt to lighting, can deal with very dark and very bright conditions, have an amazing field of view, can accurately track tiny objects very far away, and very quickly refocus and swivel our eyes and head. You'll have a lot of trouble finding cameras+mounts that meet all of these criteria that aren't a giant bullseye. It's more likely that you need to tailor your sensors for particular situations. What we see as a person at 80 meters might be a few pixels if the camera is zoomed out. If you zoom in, you'll see the enemy, but not your surroundings.
Communications. Quickly transmitting binary data at the ground level is a very open-ended problem.
Positioning. GPS doesn't work within buildings. Some of your platform power will be allocated to dealing with this fact.
http://wiki.lesswrong.com/wiki/Paperclip_maximizer
A slight mis-prioritization on the AI's goals could easily lead to spectacularly unintended results.
Ok, so rather than 'winning' it's achieving the desired change through application of force. If the point of warfare was inflicting enemy casualties, than we would have already won in Iraq and Afghanistan... and Vietnam. We've been moving away from the whole 'kill as much people as possible' method of warfare for quite a while.
Omelets require breaking eggs, but the point of omelets isn't to break eggs.