# 论文草稿：实验

> 本页由仓库源文件生成。请不要直接编辑本页；修改下方 source 文件后运行 `vp run docs:generate`。

> 中文支持说明：本页是中文站点镜像，保留论文标题、证据状态和来源字段的原始表述，避免未经核实的翻译改变证据边界。

> 来源: `paper/sections/04-experiments.md`

## Experiments

Status: `pre-data experiment design draft`

The experiment section should define the validation ladder for `SAH-BRI-Grasp`. It should not include results or statistical claims until data are collected.

## Experiment 1: SSVEP-YOLO Dynamic Target Selection

Goal: test whether scene candidates generated by an object detector can serve as reliable SSVEP targets.

Conditions:

| Condition | Purpose |
| --- | --- |
| Fixed UI button SSVEP baseline | establish controlled target-selection baseline |
| Static scene candidate boxes | test object-box layout without detector jitter |
| Recorded YOLO video candidate boxes | test realistic visual dynamics with replay control |
| Real-time YOLO candidate boxes | test full online candidate generation |

Variables:

* object count: 2, 4, 6;
* box size: small, medium, large;
* box spacing: far, near, partially overlapping;
* EEG window: 0.5 s, 1.0 s, 1.5 s, 2.0 s;
* channel subsets: occipital/parieto-occipital sets and low-channel subsets.

Metrics:

* SSVEP accuracy;
* ITR;
* response time;
* false selection rate;
* candidate jitter;
* target lost rate;
* correction count;
* fatigue and NASA-TLX.

Evidence target: update SSVEP role and dynamic command-space rows in `library/EVIDENCE_MATRIX.md`.

## Experiment 2: MI Mode Control

Goal: test whether MI can provide intervention or mode control for shared autonomy.

Command progression:

| Stage | Commands | Purpose |
| --- | --- | --- |
| Binary | intent vs rest, or confirm vs cancel | establish baseline separability |
| Ternary | confirm / cancel / rest | add no-control rejection |
| Four-state | enter-control / confirm / pause-stop / cancel | approximate full system mode control |

Metrics:

* classification accuracy;
* command latency;
* false activation rate;
* mode-switch success rate;
* training time;
* no-control rejection performance.

Evidence target: update MI role and shared-autonomy rows.

## Experiment 3: Closed-Loop Robotic Grasping

Goal: evaluate the full EEG-vision-robot loop from scene object detection to grasp execution.

Validation ladder:

1. simulation closed loop;
2. fake target or offline EEG replay loop;
3. online human EEG with physical robot.

Baseline families:

| Baseline | Rationale |
| --- | --- |
| Fixed command space vs scene-aware command space | tests dynamic command-space contribution |
| SSVEP-only vs SSVEP+MI hybrid confirmation | tests hybrid BCI contribution |
| Shared-autonomy high-level intent vs low-level discrete direction commands | tests autonomy delegation |
| Hindsight optimization / policy blending variants | ties experiment to `SA-001` and `SA-002` |

Metrics:

* task success rate;
* grasp success rate;
* completion time;
* correction count;
* false activation;
* emergency stop frequency;
* abort rate;
* target lost rate;
* planning failure rate;
* unsafe command blocked count;
* NASA-TLX;
* user preference.

Evidence target: update core hypothesis, shared autonomy, dynamic command-space, and robot grasping rows.

## Experiment 4: Low-Channel Ablation

Goal: evaluate whether the system can move from standard EEG cap acquisition toward lower-channel wearable configurations.

Strategy:

1. collect full-core 16-channel data;
2. replay 12-channel, 8-channel, 6-channel, and 3-channel subsets offline;
3. move only stable subsets into online tests.

Channel sets:

```text
16ch full-core
12ch product-core
8ch minimal-core
6ch hybrid-minimal or SSVEP-prior
3ch SSVEP-only: O1/Oz/O2
3ch MI-only: C3/Cz/C4
```

Metrics:

* SSVEP accuracy and ITR by channel subset;
* MI accuracy and false activation by channel subset;
* closed-loop task success and completion time;
* setup time and user burden.

Evidence target: update low-channel translation row.

## Analysis Plan

The baseline and analysis draft is now defined in `experiments/analysis-plan.md`. It fixes the planned baseline families, primary endpoints, common logging fields, failure-mode reporting, attempted-trial analysis unit, ITR timing variants, and advancement gates for Exp1-Exp4.

Pre-data freeze files are now available:

* `experiments/pre-data-freeze-checklist.md`;
* `experiments/hardware-inventory.md`;
* `experiments/exp03-closed-loop-grasping/hand-eye-calibration.md`;
* `experiments/templates/session-record.md`;
* `experiments/templates/threshold-freeze-record.md`;
* `experiments/templates/trial-log-schema.md`.

Still `needs confirmation` before data collection:

* participant inclusion/exclusion criteria;
* trial counts per condition;
* exact EEG montage;
* decoder training and validation splits;
* exact raw-window and practical-selection `T` definitions for ITR;
* statistical model or paired-comparison strategy;
* numeric advancement and safety thresholds;
* hardware, robot, camera, calibration target, and control-stack identifiers.

## Blocked Content

Do not add:

* participant outcomes;
* p-values or effect sizes;
* claimed task success improvements;
* claimed workload reductions;
* hardware-specific timing claims.