From an open-loop decoder to a sensorimotor loop
In Volume I you built a read-only pipeline: neural activity in, intended movement out. That is a BCI running open-loop — the brain has no direct channel back. Yet every fluent biological movement is closed-loop. As your hand closes on a cup, thousands of tactile and proprioceptive afferents report contact, force and slip within tens of milliseconds, and motor cortex uses that stream continuously. A neuroprosthesis that only reads is missing half of what the nervous system actually uses to act.
A bidirectional BCI adds the return path: it both decodes intention and encodes artificial feedback delivered by stimulation. When the read side, the effector and the write side all run in real time, you have a closed-loop sensorimotor BCI — the subject of this whole track.
Why vision cannot replace touch
Read-only motor BCIs lean entirely on vision for feedback, and vision is a poor substitute for somatosensation. It is comparatively slow and serial, it consumes scarce attention, and — critically — it cannot report contact force or the onset of slip. Users of visual-only decoders grasp cautiously, watch every object, and still crush or drop things. Somatosensory feedback is fast, parallel, and carries exactly the force/slip signals that grip control depends on.
The control-theoretic argument
The cleanest way to see why feedback is mandatory is optimal feedback control (OFC). Model the motor system as a controller that minimizes a cost balancing tracking error against effort. To act well it must maintain a running estimate \hat{x}_t of the plant state, and that estimate degrades without measurements. In a BCI, both the plant (a cursor or arm) and the decoder inject noise; only a feedback channel lets the controller correct for it.
J = \mathbb{E}\!\left[\sum_{t}\, (x_t - x_t^{*})^{\top} Q\,(x_t - x_t^{*}) + u_t^{\top} R\, u_t \right]The OFC cost: penalize deviation from the goal x_t^{*} and control effort u_t. Minimizing it requires an ongoing state estimate, which is impossible to keep accurate without feedback.
This is the scorecard the brain is trying to make as small as possible: over the whole movement it adds up how far you are from the goal plus how much effort you spent. You literally cannot minimize it without knowing where you are right now — and knowing that requires feedback.
- J
- The total cost to be minimized — smaller is better control.
- x_t - x_t^{*}
- The gap at time t between where you actually are and the goal x_t^{*}.
- Q
- Weight on how badly a state error hurts — big Q means 'stay on target'.
- u_t
- The control command (effort) issued at time t.
- R
- Weight on how costly effort is — big R means move gently and cheaply.
- \mathbb{E}
- Expected value — the average once you account for noise and randomness.
Crank Q up and the user hurries straight at the target; crank R up and they glide there slowly to save effort.
Feedback also comes with a deadline. Delivered too late, it can destabilize rather than help — a controller acting on stale state can overshoot and oscillate. Every millisecond of loop latency therefore matters, which is why Guide 4 treats the latency budget as a first-class design constraint.
Reading and writing at once
Concretely, a somatosensory bidirectional BCI reads from motor cortex to drive an effector and writes into somatosensory cortex (areas 3b/1) with intracortical microstimulation (ICMS) to evoke touch. The catch is that the two directions share hardware and tissue: you must record microvolt spikes while injecting stimulation pulses that are orders of magnitude larger. That coexistence — read–write interleaving — is the central engineering problem of the field.
- Read: acquire motor-cortex activity and decode the intended effector command.
- Act: move the cursor / arm and register the resulting contact, force and slip.
- Encode: map that contact into a stimulation pattern (amplitude, frequency, electrode).
- Write: deliver ICMS to somatosensory cortex, evoking an artificial percept.
- Adjust: the user perceives the feedback and updates the next command — the loop repeats.