Latency
Bots get commands, decrypt them, and then move, make telemetry and image data, encrypt them, and send the parcel back. Data is decrypted, then a human decides what to do next. As a result, all data from the frontline is at least 5 to 30 seconds old. By the time the operator sees the tank coming, the drone might be dead already, would it engage in hostilities. So, why arm up a drone with equipment to lose if your latency will lead to its loss? Ordering an attack to a drone with that much latency is suicidal.
So the autonomous drone is set up to dodge, hide, and evade on its own, and on a secondary level follow the location instructions by the human operators in the control bunker. Even if the imagery is still somewhat time-delayed, it is good enough to plan larger operations.
Prevent Robot Uprising
Arming up bots to the teeth does not just make them large and bulky, it makes people uneasy because they fear robot uprisings to murder humans. Teaching an AI to kill some humans and then expecting them to not kill some of them sounds terribly insecure to the general population.