# 综合主线 > 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: `library/SECONDARY_SYNTHESIS_MAINLINES.md` ## Secondary Synthesis Mainlines This file records synthesis hypotheses. Do not promote them to final claims until linked paper cards and evidence matrix rows exist. ## Mainline 1: EEG As High-Level Intention Status: `inferred` The project hypothesis is that non-invasive EEG should not carry low-level continuous control for robotic grasping. Instead, SSVEP and MI should express high-level user intention, while robot autonomy handles planning and execution. Current evidence: * `BCI-001` supports BCI as a limited-capacity, adaptive communication/control channel and recommends matching applications to users and signal capacity. * `BCI-002` supports the need to distinguish offline decoder evidence from online BCI operation, with attention to calibration burden, limited training data, and real-life EEG noise. * `BCI-003` reinforces the BCI pipeline decomposition, SSVEP/MI control-signal tradeoffs, and the rule that online evidence is needed for solid system-performance claims. * `BCI-004` supports modular online BCI runtime design, timing/jitter measurement, and standardized event/data logging. * `BCI-005` adds user-facing constraints around training burden, sensor comfort, wearable electrode placement, visual-stimulation safety, and integration with other interfaces. * `MI-005` verifies that noninvasive EEG motor imagery can control reach/grasp tasks when decomposed into low-dimensional sequential commands, while not establishing fluid 3D continuous arm control. * `MI-006` supports early adaptive online MI control and hand-orthosis operation, while warning that feedback, co-adaptation, and asynchronous detection remain central constraints. * `MI-009` verifies noninvasive EEG control of a mobile robot through high-level mental commands combined with robot autonomy, not direct low-level motor control. * `BRI-001` supports hybrid MI-SSVEP fusion as an offline decoder baseline, while preserving the online/robot-grasp validation gap. * `BRI-002` verifies that BCI robotic manipulation can benefit from vision, intent inference, and autonomy, but uses intracortical rather than noninvasive EEG. * `BRI-003` supports behavior-tree task structuring for noisy or low-dimensional teleoperation inputs, but remains an architecture proposal without EEG robot validation. * `BRI-004` verifies a noninvasive AR BRI comparator where gaze and MI carry high-level object/action choices while robot policies execute manipulation. * `BRI-005` and `BRI-006` provide invasive direct-control upper-bound comparators; they should not be used as evidence that low-channel noninvasive EEG can reproduce intracortical manipulation performance. Evidence still needed: * Project closed-loop comparison against fixed command-space control. * User workload and perceived-control metrics for noninvasive SSVEP-MI control. ## Mainline 2: SSVEP For Scene-Semantic Selection Status: `inferred` SSVEP is positioned as a target-selection and confirmation channel. The key unresolved question is whether dynamic YOLO-generated boxes remain stable enough for reliable short-window SSVEP decoding. Current evidence: * `SSVEP-001` supports frequency tagging and attention-modulated SSVEP responses for multiple visual inputs. * `SSVEP-003` supports FBCCA as a high-speed 40-target SSVEP decoding baseline. * `SSVEP-004` supports TRCA/ensemble TRCA as a short-window, high-speed SSVEP decoder with individual calibration. * `SSVEP-005` provides a 40-target, 35-subject, 64-channel benchmark with frequency/phase coding, data-length/ITR analyses, and explicit electrode-number/location optimization potential. * `SSVEP-007` verifies that flicker responses are frequency-dependent and show resonance peaks, supporting careful frequency/display/comfort protocol design. * `SSVEP-010` supports Compact-CNN as an asynchronous SSVEP comparator, while preserving the boundary that offline fixed-keypad evidence is not dynamic object-box evidence. * `SSVEP-011`, `SSVEP-012`, and `SSVEP-013` anchor early SSVER/VEP/SSVEP BCI systems before modern CCA/FBCCA/TRCA baselines. * `SSVEP-014` provides a VEP/SSVEP system-design review that separates transient and steady-state VEPs and emphasizes practical BCI constraints. * `SSVEP-016` verifies CCA as a classical SSVEP frequency-recognition method, while `SSVEP-019` shows that calibration-informed CCA variants can improve detection over standard CCA. * `SSVEP-020` supports high SSVEP user coverage in a 53-subject four-LED study, while preserving the boundary that this is simpler than a dynamic object interface. * `SSVEP-021` and `SSVEP-029` add large public dataset evidence through BETA and Dual-Alpha. * `SSVEP-022` and `SSVEP-023` provide AR/VR SSVEP evidence, but not YOLO-bound physical-object SSVEP. * `SSVEP-026`, `SSVEP-027`, `SSVEP-028`, `SSVEP-030`, and `SSVEP-031` document recent trends in transformers, augmentation, one-shot transfer, priors, and attention-based SSVEP decoding. Evidence still needed: * Detector-bound object-level SSVEP studies beyond fixed spellers/keypads and generic AR/VR SSVEP. * Experiment 1 results. ## Mainline 3: MI For Mode Control Status: `inferred` MI is scoped to active intervention and state control rather than 6DOF robot motion. This may reduce control burden but still needs evidence around false activation, latency, and training cost. Current evidence: * `MI-002` verifies CSP as a foundational motor-imagery spatial-filter method and records sensitivity to artifacts and electrode placement. * `MI-006` verifies early MI BCI with online feedback and hand-orthosis control, while preserving calibration/co-adaptation/asynchronous-control cautions. * `MI-007` verifies FBCSP as a strong MI benchmark method, including 22-channel and 3-bipolar-channel BCI Competition IV contexts. * `MI-008` verifies BCI Competition IV context for no-control/rest periods, session transfer, artifacts, kappa evaluation, and online-delay concerns. * `MI-005` verifies noninvasive MI control of robotic reach/grasp through decomposed low-dimensional commands. * `MI-009` verifies noninvasive EEG mobile-robot control using high-level mental commands, confidence rejection, and robot finite-state autonomy. * `MI-005` also supports the limitation that high-accuracy noninvasive 3D continuous arm control was not demonstrated. * `MI-004` supports compact EEGNet decoding across P300, ERN, MRCP, and SMR, including 22-channel SMR benchmark data, but only as offline evidence. * `BRI-001` supports hybrid MI-SSVEP decoding as a two-stream CNN comparator, but uses offline 4-second epochs and does not test robot control. * `BRI-004` supports online MI as a high-level Place/Use action selector in an AR BRI, with gaze-based recovery improving effective online decisions. * `MI-013` adds the classical EEG-BCI classifier review layer before the 2018 update. * `MI-014` supports CSP as a practical spatial-filter method for robust single-trial MI/SMR analysis. * `MI-016` and `MI-027` support the Riemannian geometry branch, from the 2012 multiclass covariance-manifold method to recent Riemannian/deep EEG review context. * `MI-017` supports the nonstationarity problem through stationary CSP. * `MI-018` and `MI-023` support transfer learning as a response to subject/session shift and calibration burden. * `MI-020`, `MI-021`, `MI-022`, and `MI-024` support the deep/end-to-end/self-supervised EEG trend while preserving offline validation boundaries. * `MI-025` and `MI-028` add wearable and gamified-training productization context. * `MI-029` shows selected-subject virtual 3D MI control but should be treated as a boundary comparator, not physical robot-arm evidence. * `MI-030`, `MI-031`, and `MI-032` support clinical/rehabilitation context only as a future product direction, not current SAH-BRI-Grasp efficacy evidence. Evidence still needed: * Experiment 2 results with no-control state evaluation. * Project-specific false activation, latency, and no-control evidence. ## Mainline 4: YOLO As Dynamic Command-Space Generator Status: `inferred` YOLO is treated as more than an object detector: its outputs define the current set of selectable semantic objects. This is the system-level bridge between vision and BCI. Current evidence: * `YOLO-001` verifies real-time class, bounding-box, and confidence outputs that can seed candidate generation. * `YOLO-002` verifies YOLOv2/YOLO9000 detector evolution, including real-time VOC performance and joint detection/classification through WordTree. * `YOLO-003` verifies YOLOv3 as a low-latency detector comparator with strong AP50 performance but weaker strict COCO AP. * `YOLO-004` verifies the two-stage RPN/Faster R-CNN detector family and provides a comparator to one-stage detectors for object candidate generation. * `YOLO-005` verifies COCO as a natural-context object detection and instance-segmentation dataset, useful for detector grounding but not for BCI or grasp validation. * `YOLO-007` verifies RetinaNet/focal loss as an accuracy-focused dense one-stage detector comparator and class-imbalance reference. * `YOLO-009` verifies YOLOv4 as a real-time detector comparator with COCO AP/AP50 and BoF/BoS training evidence. * `YOLO-010` verifies YOLOv7 as a modern real-time detector comparator with higher AP/FPS tradeoffs than older YOLO-family baselines. * `GRASP-001` verifies that object detection is not sufficient for grasp execution; grasp synthesis still needs candidate generation, ranking, reachability filtering, and robust execution. * `GRASP-003` verifies RGB-D CNN grasp-rectangle prediction as a separate task from object detection. * `GRASP-004` verifies robust depth-based grasp-candidate ranking with GQ-CNN/Dex-Net and records depth/collision failure modes. * `GRASP-005` verifies a closed-loop depth-based grasping module that predicts grasp quality, angle, and width per pixel, but still depends on camera parameters and known robot-camera calibration. * `GRASP-006` verifies the classical eye-on-hand calibration rationale: camera-frame measurements must be related to robot/gripper coordinates before physical grasp execution. * `GRASP-007` verifies large-scale self-supervised learned hand-eye coordination and continuous visual servoing, while preserving data-scale, hardware-transfer, occlusion, and depth-related limitations. Evidence still needed: * Project-specific grasp-pose implementation choice. * Project-specific hand-eye calibration implementation and acceptance criteria. * Safety evidence for false detection, target loss, and confirmation gates. ## Mainline 5: Closed-Loop Shared Autonomy Status: `inferred` The proposed system should be evaluated as an EEG-vision-robot loop. The strongest paper claims should depend on task success rate, grasp success rate, completion time, false activation, unsafe command blocking, and user workload. Current evidence: * `SA-001` verifies hindsight-optimization shared autonomy can reduce execution time and user input in a robot manipulation user study, while preference/control concerns remain. * `SA-002` verifies policy blending and confidence-dependent arbitration principles, including the risk of aggressive assistance under wrong intent prediction. * `MI-009` verifies a noninvasive robot-control example where high-level EEG commands are interpreted by robot autonomy rather than directly mapped to motor outputs. * `BRI-002` verifies BCI manipulation benefits from vision and shared autonomy, with the caveat that the BCI is intracortical. * `BRI-003` supports behavior-tree representations for assistive teleoperation phases, while leaving empirical validation open. * `BRI-004` verifies a noninvasive AR BRI feasibility study with 18 healthy users, high subtask success, online EEG action decoding, and SUS/NASA-TLX reporting. * `BRI-005` and `BRI-006` provide invasive direct-control manipulation comparators for reach, grasp, and high-dimensional prosthetic-arm control. * `BCI-004` supports the need for explicit runtime modules, event markers, timing logs, and standardized data records in online BCI systems. * `GRASP-005` verifies robot-side closed-loop grasp recovery under dynamic objects and clutter, supporting the autonomy side of the loop. * `GRASP-004` adds a robot-side failure taxonomy around missing depth geometry and collision misclassification that should be preserved in Exp3 logs. * `GRASP-007` adds robot-side learned visual servoing evidence and shows why local hardware/data transfer limits and repeated-failure logs matter. Evidence still needed: * Experiment 3 and Experiment 4 results. * Explicit baseline comparisons. * Failure-mode and recovery logs.