# Motor Imagery Literature

> Generated from repository source files. Do not edit this page directly; edit the source file listed below and rerun `vp run docs:generate`.

> Source: `literature/02-mi-history-and-paradigms.md`

## MI History And Paradigms

Status: `inferred`

This page summarizes the local motor-imagery evidence after the 2026-07-12 deep-dive intake. It supports MI as an active mode/intervention channel, but it does not verify safe online stop/cancel control in SAH-BRI-Grasp.

## Review Questions

* What ERD/ERS and sensorimotor rhythm evidence supports motor imagery decoding?
* How did MI decoding move from CSP to FBCSP, Riemannian methods, transfer learning, and deep/self-supervised learning?
* What do training burden, nonstationarity, and BCI inefficiency imply for product design?
* What is the boundary between virtual continuous MI control and physical robotic grasping?
* How should rehabilitation evidence be separated from the first engineering system paper?

## Local Evidence Map

| Evidence Need | Candidate IDs | Status |
| --- | --- | --- |
| MI foundation and early direct BCI | MI-006 | extracted |
| CSP/FBCSP baseline | MI-002; MI-007; MI-014 | extracted |
| BCI Competition context | MI-008 | extracted |
| Classical classifier review | MI-013 | extracted |
| Nonstationarity | MI-017 | extracted |
| Riemannian geometry | MI-016; MI-027 | extracted |
| Transfer learning / calibration reduction | MI-018; MI-023 | extracted |
| Deep/end-to-end/self-supervised EEG | MI-004; MI-020; MI-021; MI-022; MI-024 | extracted |
| Wearable/productization and training UX | MI-025; MI-028 | extracted |
| Robot/continuous-control comparators | MI-005; MI-009; MI-029; BRI-004 | extracted |
| Clinical/rehabilitation direction | MI-030; MI-031; MI-032 | extracted |
| Project no-control and false activation | project experiment needed | needs confirmation |

## Lineage

### 1. Sensorimotor Rhythm And Classical Spatial Filtering

`MI-006` anchors the early MI BCI line. `MI-002`, `MI-014`, and `MI-007` then provide the CSP/FBCSP family: spatial filtering is central because scalp EEG mixes sources, and subject-specific frequency/spatial patterns matter.

`MI-013` supplies the older classifier taxonomy, while `MI-008` places MI decoding in the BCI Competition IV benchmark context with rest/no-control, session-transfer, artifacts, and kappa-style evaluation.

### 2. Nonstationarity, Riemannian Geometry, And Transfer

The added papers make the MI difficulty clearer:

* `MI-017` shows why nonstationarity challenges ordinary CSP.
* `MI-016` introduces Riemannian covariance geometry as an alternative to a separate spatial-filter-plus-classifier pipeline.
* `MI-027` shows that Riemannian methods remain a current topic, including deep/Riemannian integration.
* `MI-018` and `MI-023` frame transfer learning as a response to subject/session shift and calibration burden.

For SAH-BRI-Grasp, this means Exp2 should not only report offline accuracy. It should test rest/no-control false activation, cross-session drift, calibration time, and latency.

### 3. Deep And Self-Supervised MI/EEG Models

`MI-004`, `MI-020`, and `MI-021` support compact/end-to-end neural decoding baselines. `MI-022` adds broader systematic-review caution: deep learning can improve EEG analysis, but datasets are heterogeneous and generalization is not automatic. `MI-024` adds the self-supervised EEG representation direction.

These are trend and comparator papers. They do not eliminate the need for CSP/FBCSP/Riemannian baselines in the first experiment.

### 4. Training Burden, Wearable Translation, And Productization

`MI-025` shows wearable MI-BCI as a product-translation direction, but wearable evidence is not the same as proof for the project's target headset or channel montage.

`MI-028` is important because it treats MI-BCI training as a user-experience problem. Gamification and training design matter because MI users often need lengthy training and a nontrivial fraction of users may struggle to reach reliable control.

### 5. Robot And Continuous-Control Boundary

`MI-005` and `MI-009` support the project's high-level-control interpretation: noninvasive EEG robot control is more credible when commands are decomposed or fused with autonomy.

`MI-029` is a useful boundary paper: selected trained subjects controlled a virtual helicopter in 3D using MI/SMR, but this was virtual, small-sample, and training-heavy. It should not be used as evidence that noninvasive MI can safely drive a physical 6-DoF arm during grasping.

### 6. Rehabilitation Direction

`MI-030`, `MI-031`, and `MI-032` support the existence of clinical and rehabilitation evidence around MI/BCI/robotics. For this repository, they should feed the rehabilitation product direction, not the main engineering-system claim, unless a separate clinical protocol is created.

## Implications For SAH-BRI-Grasp

* Treat MI as `mode`, `confirm`, `pause`, `cancel`, or `intervention`, not low-level continuous arm control.
* In Exp2, include rest/no-control states and report false activation, latency, and calibration burden.
* Use CSP/FBCSP/Riemannian as transparent baselines before deep models.
* Treat wearable and gamified training papers as productization evidence, not immediate system validation.
* Keep rehabilitation claims separate from the first scene-aware grasping paper.

## Remaining Gaps

| Gap | Current Handling |
| --- | --- |
| No project-specific no-control evidence. | Exp2 must include rest/no-control trials and false activation metrics. |
| No local online MI latency evidence. | Measure command delay and safety-gate timing in the prototype loop. |
| Low-channel MI performance is untested. | Run Exp4 channel ablation after collecting full-channel data. |
| Rehabilitation evidence exists but is clinically separate. | Do not claim therapeutic benefit without a dedicated clinical design. |