Can't the brain alone use visual feedback to close the loop, separating both hands? If thinking in a particular way activates the only the robot hand (trembling), I'd guess that with sufficient plasticity you should be able to start enhancing control of it.
Obviously tactile and proprioceptive feedback through the chip would enable better / quicker control, given the sufficient plasticity hypothesis.
I suspect it would be possible if you can do some kind of real-time feedback/tuning of the ML model. Reinforcement learning stuff.
By slowly adjusting the model to reward hitting the ball back in pong, but also to penalize moving the real physical hand, you might achieve it.