# 论文草稿：讨论

> 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/06-discussion.md`

## Discussion

Status: `limitation scaffold drafted`

This section is a claim-gated discussion scaffold. It can describe expected interpretation axes and limitations, but it must not state experimental findings before results exist.

## Interpretation Boundary

`SAH-BRI-Grasp` should be discussed as an evidence-motivated architecture until local experiments are complete. The current literature supports the plausibility of the component roles, but not the final closed-loop performance of this specific system.

Safe discussion claims:

* noninvasive EEG is better scoped to high-level intent, confirmation, mode control, and supervision than to dense low-level arm motion;
* SSVEP frequency-tagging and FBCCA/TRCA-style decoders provide a plausible basis for target selection;
* MI can be framed as a low-bandwidth supervisory channel rather than continuous end-effector control;
* object detection can generate candidate command spaces, but grasp execution requires grasp-pose estimation, calibration, planning, and safety checks;
* shared autonomy can be used to arbitrate between uncertain human intent and robot-side feasibility.

Blocked discussion claims:

* improved task success;
* reduced workload;
* faster completion;
* reliable low-channel operation;
* validated physical hand-eye accuracy;
* superiority over a baseline.

## Literature Transfer Limits

### Fixed SSVEP Spellers To Scene Objects

`SSVEP-003`, `SSVEP-004`, and `SSVEP-005` support high-speed SSVEP decoding in controlled fixed-target settings. `SAH-BRI-Grasp` transfers that idea to object-bound scene candidates. The transfer is not automatic because object boxes may move, disappear, overlap, vary in size, or compete with scene clutter. Exp1 must therefore decide whether YOLO-generated or frozen scene candidates can preserve usable SSVEP performance.

### MI Robotic Control To Mode-Level Supervision

`MI-005` supports noninvasive EEG control for reach-and-grasp tasks, but the project does not assume continuous high-degree-of-freedom arm control from EEG. The defensible interpretation is narrower: MI is a candidate mode-control and intervention channel. Exp2 must test false activation and no-control rejection before MI can be used as a safety-relevant online command.

### Object Detection To Grasp Execution

`YOLO-001` supports object detection as a real-time visual perception primitive, while `GRASP-001` and `GRASP-005` separate detection from grasp-pose generation and closed-loop execution. `GRASP-006` supports the classical need for hand-eye calibration in eye-on-hand robotic systems. Therefore, a positive detector result would not by itself validate grasping; Exp3 must evaluate candidate tracking, transform consistency, planning, grasping, and verification.

### Shared Autonomy Formalisms To Hybrid BCI

`SA-001` and `SA-002` provide shared-autonomy formalisms, but not evidence that noninvasive SSVEP-MI intent can safely drive the proposed grasping pipeline. The discussion should treat shared autonomy as the robot-side mechanism for handling uncertain, low-bandwidth intent, not as proof of project performance.

## Failure Modes To Discuss

The discussion should report failure modes even if aggregate metrics look acceptable:

| Failure Mode | Interpretation |
| --- | --- |
| target lost during SSVEP window | scene-aware command space may be too unstable |
| wrong SSVEP selection | stimulus binding or decoder confidence may be insufficient |
| MI false activation | mode-control channel may be unsafe without stronger no-control rejection |
| conflicting SSVEP and MI evidence | fusion policy may need conservative arbitration |
| grasp-pose failure | detector output is not enough for manipulation |
| calibration or transform error | hand-eye procedure may not support physical execution |
| planning failure | selected object may be unreachable or unsafe |
| safety gate over-triggering | autonomy may be too conservative for practical use |

## How Results Should Be Interpreted

Positive results would support different claims depending on which experiment succeeds:

| Evidence Source | Claim It Could Support | Claim It Cannot Support Alone |
| --- | --- | --- |
| Exp1 | scene candidates can function as SSVEP targets under tested conditions | closed-loop grasp success |
| Exp2 | MI can provide mode-level intervention under tested conditions | safe robot execution |
| Exp3 simulation/replay | software integration and baseline comparison under controlled conditions | physical robot reliability |
| Exp3 physical robot | closed-loop feasibility for the tested hardware and object set | broad deployment readiness |
| Exp4 offline ablation | candidate low-channel subsets | wearable or online low-channel performance |
| Exp4 online test | online feasibility for selected low-channel subsets | general low-channel robustness |

## What Would Refute The Design

The discussion should explicitly handle negative outcomes:

* if live scene candidates sharply reduce SSVEP accuracy, the system may need candidate freezing, larger targets, fewer simultaneous objects, AR-style overlays, or non-SSVEP target selection;
* if no-control false activation remains high, MI should be limited to non-safety-critical confirmation or removed from online robot execution;
* if fixed command spaces outperform scene-aware command spaces, the visual perception module may add complexity without user benefit;
* if calibration or grasp-pose errors dominate, the contribution should be reframed around interface design rather than physical grasp execution;
* if low-channel subsets degrade strongly, the wearable path should remain future work.

## Final Discussion Blockers

* closed-loop experiment results;
* hardware timing and calibration evidence;
* user workload and preference results;
* participant-level variability analysis;
* baseline statistical comparisons;
* failure-mode counts from the trial logs.