# 论文草稿：相关工作

> 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/02-related-work.md`

## Related Work

Status: `draft ready`

This draft is evidence-gated. Component-level findings are drawn from extracted paper cards; claims about the integrated `SAH-BRI-Grasp` system remain design rationale until project experiments are completed.

## BCI As High-Level Intention Interface

Brain-computer interfaces are commonly framed as non-muscular communication and control channels that translate electrophysiological activity into device commands (`BCI-001`). This framing is important for robotic manipulation because EEG-based control is not simply a replacement for a joystick or a low-level arm controller. The BCI loop depends on user adaptation, feedback, feature extraction, translation algorithms, signal-to-noise limits, and application matching. `BCI-001` also distinguishes user-level signal control from system-level task performance, which motivates evaluating `SAH-BRI-Grasp` at both levels rather than reporting decoder accuracy alone.

For the present system, the literature supports a conservative control allocation: noninvasive EEG should express high-level intent, while the robot handles low-level movement, planning, and grasp execution. This is a design inference, not yet a verified result for `SAH-BRI-Grasp`. It follows from the limited-capacity BCI framing in `BCI-001`, the decomposed noninvasive motor-imagery control demonstrated in `MI-005`, and BCI manipulation systems in which vision and autonomy improve physical task execution (`BRI-002`, `BRI-004`).

The remaining BCI foundation evidence adds implementation and usability constraints. `BCI-002` shows that classifier evidence must be handled carefully because many EEG classification methods are evaluated offline, while practical BCI operation is online and sensitive to calibration time, noise, and limited training data. `BCI-003` reinforces the BCI pipeline decomposition and the difference between exogenous SSVEP-like selection channels and endogenous MI-like control channels. `BCI-004` provides a modular online BCI runtime model with source, signal-processing, application, and operator modules, and evaluates latency/jitter explicitly. `BCI-005` adds a user-facing perspective from BCI games: training burden, sensor comfort, visual-stimulation safety, and integration with other interfaces affect adoption. These papers support the current method choice to keep decoder outputs, stimulus events, runtime records, robot events, and user-experience measures separated in the `SAH-BRI-Grasp` interface skeleton.

## SSVEP Target Selection And Benchmark Decoding

SSVEP is a natural candidate for target selection because periodic visual stimulation produces frequency-specific EEG responses, and multiple simultaneous visual inputs can be separated through frequency tagging (`SSVEP-001`). This mechanism supports the idea of assigning visual codes to scene candidates. However, the supporting evidence is strongest for controlled visual layouts, not for moving object boxes generated by an object detector.

Several SSVEP studies provide decoder and benchmark references for this module. `SSVEP-003` evaluates FBCCA in a 40-target speller and reports high online performance using nine occipital/parieto-occipital channels. `SSVEP-004` evaluates TRCA and ensemble TRCA, showing that calibrated spatial filtering can improve short-window high-speed SSVEP decoding in fixed speller conditions, while also introducing calibration and gaze-shift burdens. `SSVEP-005` provides a 35-subject, 64-channel, 40-target benchmark dataset with frequency/phase coding and time-window/ITR analyses. This benchmark is especially relevant for offline channel-ablation and target-window planning.

Two additional SSVEP sources sharpen the experiment boundary. `SSVEP-007` shows broad flicker responses and resonance peaks around specific frequencies, which means Exp1 frequency choices should be treated as protocol variables tied to display hardware, comfort, and safety screening. `SSVEP-010` evaluates a compact CNN for asynchronous SSVEP classification without user-specific calibration on a fixed keypad dataset, making it useful as an asynchronous decoder comparator. However, it still does not validate moving or detector-generated object boxes.

The unresolved transfer is the core SSVEP problem for `SAH-BRI-Grasp`: fixed character-grid spellers are not the same as detected, jittering, partially occluded, or task-dependent scene objects. Therefore, this paper should cite SSVEP work for frequency tagging, decoder baselines, and channel/time-window planning, while treating dynamic object-box SSVEP selection as an experiment target.

## Motor Imagery For Mode-Level Control

Motor imagery has been used for noninvasive robotic-arm control, but the strongest local evidence supports decomposed and low-dimensional control rather than dense continuous 6DOF arm operation. `MI-005` reports noninvasive EEG motor-imagery control of a JACO robotic arm for reach-and-grasp tasks, using staged cursor/arm control, hover confirmation, and low-dimensional command decomposition. The same card records an important limitation: the study does not establish fluid high-accuracy 3D continuous control from multiple independent noninvasive EEG signals.

This boundary fits the proposed `SAH-BRI-Grasp` role for MI. MI should be scoped to mode-level decisions such as confirm, cancel, pause, stop, or execute, while the robot performs low-level grasp planning and motion. `MI-004` adds a decoder-oriented reference: EEGNet is a compact CNN validated across several BCI paradigms, including SMR data from BCI Competition IV Dataset 2A, and uses depthwise and separable convolutions to reduce parameter counts. This supports EEGNet as an offline baseline or ablation model, but it does not verify online no-control behavior, false activation handling, or robot task performance.

Classical MI decoding evidence supports the baseline set for Exp2. `MI-002` verifies common spatial patterns as a motor-imagery spatial-filtering method and also records practical limitations around artifact-free data and electrode-position stability. `MI-007` extends this line with filter-bank CSP on BCI Competition IV datasets, including a 22-channel four-class dataset and a 3-bipolar-channel two-class dataset. Together, these papers justify CSP/FBCSP baselines and low-channel MI ablations, but they do not remove the need to test no-control false activation in the project protocol.

Two additional MI sources sharpen this control boundary. `MI-006` reports early motor-imagery BCI work with online feedback and hand-orthosis control, but also emphasizes co-adaptation, feedback effects, and the unresolved difficulty of asynchronous purely mental-event detection. `MI-008` reviews BCI Competition IV and records no-control/rest periods, session-transfer, artifact, kappa, and online-delay constraints. These sources support treating MI as a calibrated intervention channel whose false activations and delay must be measured inside the robot loop.

`BRI-004` provides a recent noninvasive comparator in which MI is used for high-level action selection rather than low-level robot motion. Its AR BRI maps MI to Place/Use commands after gaze-based object selection and reports both offline and online decoder metrics. The input split differs from `SAH-BRI-Grasp`, but the system-level role of MI is aligned: MI expresses a compact action or intervention signal inside a shared-autonomy loop.

`MI-009` provides an older but useful noninvasive robot-control comparator. It maps EEG-recognized mental states to high-level mobile-robot commands and lets robot autonomy handle wall following, turning, obstacle avoidance, and stopping. This supports the same design principle as `SAH-BRI-Grasp`: the BCI should modulate a robot controller rather than directly drive every actuator. `BRI-001` adds a hybrid MI-SSVEP decoder baseline through a two-stream CNN, but it remains offline, binary, and based on 4-second epochs, so it should be used as a decoder comparator rather than system evidence.

## Visual Perception And Dynamic Command Spaces

Scene-aware BRI requires converting the physical scene into a set of selectable user intentions. Object detection is one route to this conversion. `YOLO-001` verifies that a one-stage detector can output object classes, bounding boxes, and confidence scores at real-time rates. These outputs are not robot commands by themselves, but they can serve as primitives for a candidate manager: class labels suggest semantic targets, bounding boxes define visual locations, and confidence scores support filtering and ranking.

`YOLO-004` provides the complementary two-stage detector family through Faster R-CNN and RPN. It verifies region proposal and detector-feature sharing, multi-scale/multi-aspect anchors, and benchmark object detection performance. For `SAH-BRI-Grasp`, this matters less as a specific implementation choice and more as a comparator class: one-stage YOLO-style detectors favor high-rate candidate generation, while two-stage detectors provide a different accuracy/speed tradeoff and proposal structure.

The detector evidence now spans a wider speed/accuracy spectrum. `YOLO-002`, `YOLO-003`, `YOLO-009`, and `YOLO-010` trace the YOLO-family development from YOLOv2/YOLO9000 through YOLOv3, YOLOv4, and YOLOv7, supporting real-time candidate-generation comparisons under different accuracy and latency targets. `YOLO-005` adds COCO as natural-context detection and instance-segmentation dataset context. `YOLO-007` adds RetinaNet and focal loss as an accuracy-focused dense-detector comparator. These papers help define detector baselines and logging variables, but none studies BCI, SSVEP stimulus binding, target freezing, grasp planning, or user confirmation.

The project-level claim that detected objects form a dynamic command space remains an inference. The related-work position should therefore be precise: vision generates candidate objects; the BCI and shared-autonomy layers decide whether, when, and how those candidates become executable commands.

## From Object Detection To Grasp Execution

Robotic grasping requires more than selecting an object box. `GRASP-001` surveys data-driven grasp synthesis and supports separating object detection or pose estimation from grasp-hypothesis generation, ranking, reachability filtering, and robust execution. This is the key boundary for `SAH-BRI-Grasp`: a YOLO-like detector can identify what the user may want, but a separate robot-side module must decide how to grasp it.

`GRASP-003` provides an RGB-D CNN grasp-detection reference that predicts oriented grasp rectangles rather than object boxes. It reinforces the distinction between semantic object detection and grasp-pose detection. `GRASP-004` adds a robust grasp-planning reference through Dex-Net 2.0 and GQ-CNN, where grasp candidates are ranked from depth data and physical failures include missing depth geometry and collision misclassification. These sources support treating grasp pose, depth quality, collision pruning, and verification as robot-side evidence streams separate from BCI selection.

`GRASP-005` provides a concrete closed-loop grasping reference. Its GG-CNN predicts grasp quality, angle, and gripper width at each depth-image pixel, enabling real-time closed-loop grasp synthesis from RGB-D depth input. The card records a 19 ms pipeline on the reported GPU desktop, closed-loop grasp generation up to 50 Hz, and real-robot experiments with dynamic objects and dynamic clutter. These findings support a low-level autonomy layer that can react during execution, which is consistent with using EEG only for high-level intent.

`GRASP-007` provides a complementary learned hand-eye coordination reference. It uses large-scale self-supervised grasp attempts and continuous visual servoing to predict grasp success from monocular images and task-space commands. This supports the broader robot-side autonomy argument, but its data-scale, hardware-transfer, occlusion, lack-of-depth, and repeated-failure limitations mean it should be cited as a grasping/control reference rather than as a ready substitute for project-specific grasp validation.

At the same time, `GRASP-005` preserves an important dependency: image-space grasps are converted to world coordinates using camera intrinsics and known robot-camera calibration. `GRASP-006` provides the classical eye-on-hand calibration rationale by estimating the relative 3D pose between a camera and robot gripper. Any method section can cite this as calibration background, but the project must still define its own hardware-specific solver, station plan, calibration target, and acceptance metric.

## Shared Autonomy And Arbitration

Shared autonomy provides the formal bridge between uncertain human intent and robot execution. `SA-001` formulates shared autonomy under uncertain user goals and shows that assisting over a goal distribution can reduce execution time and user input in a robot manipulation user study. It also shows that better task performance does not automatically imply stronger user preference, making perceived control and preference relevant evaluation outcomes.

`SA-002` gives a complementary policy-blending formalism. It shows that arbitration should depend on confidence in the predicted goal and that aggressive assistance can be harmful when intent prediction is wrong. This is directly relevant to BCI-driven manipulation because EEG decisions can be uncertain, delayed, or wrong. For `SAH-BRI-Grasp`, shared autonomy should therefore be conservative until object detection, SSVEP selection, MI state, and grasp feasibility agree strongly enough to proceed.

These shared-control studies do not use EEG. Their value is methodological: they define baseline formalisms and failure risks for assistance under uncertain intent. The proposed experiments should therefore compare scene-aware command generation against fixed command spaces and include cancel, pause, recovery, and confirmation paths rather than assuming that autonomy should always act aggressively.

`BRI-003` adds a system-architecture comparator through behavior-tree assisted teleoperation. It supports phase-based task structure for noisy or low-dimensional user inputs, but the local evidence does not include empirical users, an EEG protocol, a detector, or a validated robot platform. It is therefore useful for designing approach, grasp, lift, recovery, and cancellation phases, not for claiming BCI-grasping performance.

## BRI Manipulation With Vision And Autonomy

`BRI-002` is the strongest local precedent for BCI manipulation with vision and shared autonomy. It combines intracortical BCI, object perception, intent inference, arbitration, and robot manipulation, and reports large improvements over direct low-level BCI control in adapted manipulation tasks. Its limitation is equally important: the BCI input is intracortical, object models and grasp poses were predefined, and transfer to noninvasive SSVEP-MI remains unverified.

`BRI-004` is a closer noninvasive comparator. It combines AR feedback, gaze-based object selection, MI-based Place/Use action selection, and shared-autonomy robot execution in multi-step ADL-like tasks with healthy participants. It also reports usability and workload measures, making it useful for shaping `SAH-BRI-Grasp` evaluation. However, its input design differs: gaze performs object targeting, MI selects actions, OWLv2 detects fixated objects, and a VLA policy executes manipulation. `SAH-BRI-Grasp` instead proposes an SSVEP-driven scene command space with MI as intervention or mode control.

`BRI-005` and `BRI-006` provide invasive upper-bound comparators for robotic reach, grasp, and high-dimensional prosthetic-arm control. They show what is possible with intracortical motor-cortex arrays and extensive calibration/training, but they should not be cited as evidence that low-channel noninvasive EEG can reproduce continuous 3D or 7D manipulation. Their value here is to clarify why `SAH-BRI-Grasp` deliberately moves low-level execution into robot autonomy.

The positioning is therefore clear. Existing BRI manipulation work supports the broad idea that BCI intent, vision, and autonomy can be combined for robot manipulation. The open question for this paper is whether noninvasive SSVEP-MI can be coupled to scene-generated object candidates in a way that remains decodable, controllable, and safe during robotic grasping.

## Evidence Boundary

| Claim Area | Current Handling |
| --- | --- |
| BCI as high-level intent | Evidence-motivated design rationale from `BCI-001`, `BCI-002`, `BCI-003`, `BCI-005`, `MI-005`, `MI-006`, `MI-009`, `BRI-001`, `BRI-002`, `BRI-003`, `BRI-004`, `BRI-005`, and `BRI-006`; not yet a `SAH-BRI-Grasp` performance result. |
| SSVEP scene target selection | Supported by SSVEP mechanism, frequency-response, fixed-speller decoder, and asynchronous comparator evidence; dynamic object-box selection remains Exp1. |
| MI mode-level control | Supported by CSP/FBCSP, compact decoder, hybrid MI-SSVEP decoder, decomposed MI robotic control, early online MI/orthosis evidence, noninvasive mobile-robot control, BCI Competition IV no-control/session-transfer evidence, and AR BRI comparator evidence; online false activation and no-control behavior remain Exp2. |
| Dynamic command space | Supported as an engineering synthesis from YOLO-family, COCO, RetinaNet, two-stage detector outputs, and BRI comparators; novelty and usability require baseline comparison. |
| Grasp execution | Supported by grasp synthesis, RGB-D grasp detection, robust grasp ranking, closed-loop depth grasping, learned visual servoing, and classical hand-eye calibration literature; project-specific grasp-pose implementation and calibration validation remain open. |
| Shared autonomy | Formal baselines are verified by `SA-001` and `SA-002`; behavior-tree teleoperation and BRI manipulation comparators add design references; transfer to noninvasive SSVEP-MI grasping remains experimental. |
| User workload and usability | Motivated by `BRI-004`; project-specific workload and perceived-control claims require local experiments. |