The idea of a human directly controlling a robot avatar in combat is an idea ultimately doomed to failure.
Humans have reaction times measured in milliseconds, while a well-programmed machine could have reaction times measured in microseconds. This would result in any mechanical avatar being shot down by any pure AI robot in short order. Adding communications lag to this setup would only make things worse.
The ideal combat automaton would be more compact than a human soldier, better armoured and not burdened with a human's abysmal reaction times.
The only realistic roles that a human might play in a battlefield filled with advanced combat robots is as the source of the objectives and rules of engagement, as a target or as an obstacle/cover (depending on the rules of engagement and if they're armed).
The other problems with humans is that they have their own agendas, including self-preservation near the top of the list. That's all very good much of the time, but sometimes in war, a sacrificial pawn is required.
In short, the best human soldier's combat expertise is of little use to a robot, especially if the soldier is directly controlling that robot. In computer first-person-shooter games, the problem that the programmers face in creating AI opponents - even ones that follow the rules and have mobility characteristics equivalent to a human - is making them imperfect enough to not win every time through sheer reactiveness, and making them smart enough is a secondary problem. Without doing this, human players get frustrated because they can't win - they stick their head out from behind cover and the 'bot kills them. They wait for the bot to stick its head out from behind cover, and the bot does so, aims and shoots before the player can react, killing a player, then retreats before any surviving players can return fire effectively. Getting more sophisticated, if players suppress an area with continuous fire, the 'bot waits for them to run out of ammo, then jumps out and returns fire between reloads - and it can count shots, and know the difference between a deliberate pause and the limits of a magazine, or it finds an alternate route, or calls in a friend.
So, don't try to make a humanoid robot that can move like a human - that's complex and a waste of computing resources. Make a small tank-like ground vehicle or an airborne drone with a minimal requirement for bodily dexterity.