Mathematical architectures for translating intracortical population vectors into multi-degree-of-freedom robotic manipulation.
Individual primary motor cortex (M1) neurons are broadly tuned to movement directions. Accurately predicting hand trajectory requires computing the weighted vector sum of firing rates across hundreds of simultaneously observed neurons.
NeuroLinkage combines steady-state Kalman filters for linear kinematic trajectory estimation with deep recurrent neural networks that model non-linear finger articulation and grasp posture transitions.
Providing tetraplegic patients with continuous visual and tactile feedback stabilizes neural decoder performance, enabling delicate tasks such as grasping a pen or picking up an egg with a robotic prosthesis.