The delay between inputs of reading every neuron, reading and actually performing a task will be the problem to solve until it opens up totally new and fastest growing sphere for the science.
Imo there are more pertinent problems than delay. When the delays are sufficiently small, say < 10mSec between neural signal and initiation of movement of the robot, which is already possible to achieve now, you get something which is really usable already and feels close enough to actually making a decision and moving your own hand for instance.
On the other hand there's longevity of electrodes, making hardware smaller etc. And then, probably biggest thing for now, the software side: developing (possibly self-learning) accurate algorithms for decoding of the signals (both single-neuron spikes + low frequency spectrum of the surrounding area) is just hard. Also because a lot of things about the brain are in essence still rather unknown it's sometimes a bit of programming in the dark.
On the other hand there's longevity of electrodes, making hardware smaller etc. And then, probably biggest thing for now, the software side: developing (possibly self-learning) accurate algorithms for decoding of the signals (both single-neuron spikes + low frequency spectrum of the surrounding area) is just hard. Also because a lot of things about the brain are in essence still rather unknown it's sometimes a bit of programming in the dark.