>I had the impression from seeing those videos of people operating robotic limbs through biofeedback or typing through biofeedback or the write-ups on that AI related work to determine words from brainwave patterns that a lot of this stuff was achieving impressive capabilities.

Sticking to BCI specifically and not other forms for biofeedback, most of those demos you see come from people with invasive, penetrative electrode arrays such as the Utah array. In some people, these sensors can enable a high level of control and performance (not Xbox controller level, but still Atari controller level which would be incredibly useful for people with Motor Neurone Disease). Unfortunately the body rejects them after a few months and they're back to square one.

To be honest with you, I don't think that the problem's intractable. Perhaps a clever materials engineer could develop a coating for the electrode arrays that the body accepts long-term and allows the user to train for years rather than months. Perhaps we could develop a new form of sensor altogether.

I don't know specifically what the answer is, but it won't come from the current paradigm of "turn garbage data into non-garbage output through the power of Machine Learning somehow" that we've been stuck in for the better part of two decades. And unfortunately, projects adhering to that paradigm is where all of the funding (and people) goes to.