Some thoughts:
Weight
The robot carry various scouting tools, mapping tools, surveillance tools etc. and these are all heavy. This means that it doesn't have the capacity to carry any extra dedicated targeting tools. Someone will probably point out that a surveillance optic and a targeting optic aren't entirely different - and to which I'd say - yes. However the surveillance optic whilst able to be used for targeting is a last-ditch effort option.
Range
Not the range of the robot - 9mm out of a PCC is effective out to about 100-200 metres - one source said that they could land shots at 400 metres (but they didn't say how reliably they were able to hit). Bearing in mind - 400 metres represents the absolute max range. Most infantry would be armed with a Rifle in a Rifle calibre - if we take the venerable 5.56 - that is effective out to 400 metres and with the same tweaking that gets a 9mm to 400 metres, the 5.56 can be taken out to almost 1 km.
Armour
See Weight - other posters have touched on this - but in order to make it practical in combat, would require more armour that would impede it's mobility.
Cost.
Depending on your Militaries view, it might be better to send humans to die than Robots. In the West (and especially the US) - the Defense budget is such that sacrificing a multi-million dollar Remote-Controlled vehicle/Drone/UAV/Thing is almost always seen as preferable to letting a Human die.
But... If you say have an excess of population, not too much money and not a whole lot of moral principals - send 10 penal combatants on a suicide mission, rather than risk an expensive and rare asset might be better.
Combat is dynamic
One of the things Humans do exceptionally well is prediction and reaction. A Computer might be able to calculate all possible moves and may even have an algorithm to help rank-order them in terms of likelihood - but Humans still have an edge for this sort of thing. For example - when I was regularly motorcycling to work (and doing extensive lane-splitting) I would observe drivers doing the 'Lane change lean' - that is, in the middle of a multi-lane highway, someone that was thinking about changing lane would turn their head to look at the lane they wanted and they would, ever so slightly, drift towards that lane. Noticing this change in behavior more than once gave me an extra 1-2 seconds of reaction time.
Additionally:
Combat algorithms are too complex
A scout algorithm of go here, look there, if threat identified - send help and if attacked point gun in general direction and scoot is relatively simple.
However, a combat algorithm of identify lanes of assault, rank order target priority, mission objectives etc. etc. before a first shot is fired and then having to constantly update all of those as things happen is too complex/taxing for the scouts CPU to handle.
