# Evidence Matrix

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

> Source: `library/EVIDENCE_MATRIX.md`

## Evidence Matrix

| Theme | Claim | Status | Supporting Paper IDs | Evidence Notes | Gaps |
| --- | --- | --- | --- | --- | --- |
| Project scope | The repository studies a scene-aware hybrid SSVEP-MI BCI for vision-guided shared-control robotic grasping. | verified | README-001 | Project README defines the direction, system flow, and four tracks. | External literature still needed for scientific framing. |
| Core hypothesis | Non-invasive EEG should act as a high-level intention interface rather than a low-level continuous robotic-arm controller. | inferred | README-001; BCI-001; BCI-002; BCI-003; BCI-005; MI-005; MI-006; MI-009; BRI-001; BRI-002; BRI-003; BRI-004; BRI-005; BRI-006 | BCI-001 supports the limited-capacity, adaptive-skill, application-matching framing; BCI-002 and BCI-003 warn that BCI classifier evidence and pipeline claims need online validation and calibration/robustness attention; BCI-005 adds user-facing training, comfort, and integration constraints; MI-005 and MI-009 show noninvasive EEG robot control through decomposed or high-level commands combined with robot behavior; MI-006 highlights co-adaptation and asynchronous-control limits; BRI-001 supports offline hybrid MI-SSVEP fusion as a decoder baseline; BRI-002, BRI-003, and BRI-004 support autonomy-assisted manipulation comparators; BRI-005 and BRI-006 are invasive upper-bound direct-control comparators rather than transfer evidence for EEG. | Need SAH-BRI-Grasp closed-loop experiment evidence and user workload evidence. |
| SSVEP role | SSVEP is suitable for semantic target selection, confirmation, cancellation, and grasp-point selection in this system. | inferred | README-001; SSVEP-001; SSVEP-003; SSVEP-004; SSVEP-005; SSVEP-007; SSVEP-010; SSVEP-011; SSVEP-012; SSVEP-013; SSVEP-014; SSVEP-016; SSVEP-019; SSVEP-020; SSVEP-021; SSVEP-022; SSVEP-023; SSVEP-026; SSVEP-027; SSVEP-028; SSVEP-029; SSVEP-030; SSVEP-031 | Added SSVEP deep-dive corpus now supports the lineage from early SSVER/VEP BCI systems, to CCA, CCA comparisons, user coverage, BETA/Dual-Alpha datasets, AR/VR feasibility, and recent transformer/GAN/one-shot/attention models. These papers strengthen the SSVEP target-selection rationale, but nearly all still use fixed targets, public datasets, or AR demos rather than detector-generated physical-object boxes. | Need dynamic-object SSVEP evidence and Exp1 target-box stability results. |
| MI role | MI is better scoped to active intervention, mode switching, confirmation, and stop/cancel than to 6DOF continuous arm control. | inferred | README-001; BCI-001; BCI-002; BCI-003; MI-002; MI-004; MI-005; MI-006; MI-007; MI-008; MI-009; MI-013; MI-014; MI-016; MI-017; MI-018; MI-020; MI-021; MI-022; MI-023; MI-024; MI-025; MI-027; MI-028; MI-029; MI-030; BRI-001; BRI-004 | Added MI deep-dive corpus now supports the lineage from classifier reviews and CSP, to stationary CSP, Riemannian geometry, transfer learning, deep/end-to-end learning, self-supervised EEG, wearable MI, gamified training, and small-sample continuous virtual control. MI-029 shows impressive selected-subject virtual 3D control, but it remains virtual and training-heavy; MI-030 supports clinical MI feasibility with user tailoring. None establish robust noninvasive 6DOF physical-arm control. | Need Exp2 false-activation, delay, and no-control results in the project loop. |
| Vision role | YOLO can generate a dynamic command space by converting scene objects into BCI-selectable candidates. | inferred | README-001; YOLO-001; YOLO-002; YOLO-003; YOLO-004; YOLO-005; YOLO-007; YOLO-009; YOLO-010 | YOLO-001 verifies real-time class, bounding-box, and confidence outputs; YOLO-002/003/009/010 support the evolution of YOLO-family real-time speed/accuracy tradeoffs; YOLO-004 and YOLO-007 provide two-stage and dense-detector comparator evidence; YOLO-005 supports natural-context object detection and segmentation dataset grounding. Treating detections as BCI commands is still a project-level engineering interpretation. | Need candidate stability, tracking/freeze, and SSVEP stimulus-binding evidence. |
| Robot grasping boundary | YOLO should be treated as object-candidate generation unless separate grasp-pose and calibration evidence supports direct grasp execution. | inferred | README-001; YOLO-001; YOLO-002; YOLO-003; YOLO-009; YOLO-010; GRASP-001; GRASP-003; GRASP-004; GRASP-005; GRASP-006; GRASP-007 | YOLO evidence verifies object detection outputs and speed/accuracy tradeoffs but not grasp poses; GRASP-001 verifies that grasp synthesis requires candidate generation, ranking, reachability filtering, and robust execution; GRASP-003 supports CNN-based grasp-rectangle prediction from RGB-D; GRASP-004 supports robust depth-based grasp ranking with GQ-CNN/Dex-Net; GRASP-005 verifies closed-loop depth-based grasp synthesis but still relies on known camera-robot calibration; GRASP-006 verifies classical eye-on-hand camera-to-gripper calibration rationale; GRASP-007 supports learned visual servoing and large-scale self-supervised grasp data while preserving hardware/data transfer limits. | Need project-specific grasp-pose implementation, hardware calibration procedure, and validation. |
| Hand-eye calibration role | Eye-on-hand robotic grasping needs a camera-to-gripper transform to relate vision measurements to robot motion. | verified | GRASP-006; GRASP-005 | GRASP-006 defines hand/eye calibration as estimating the camera pose relative to the robot gripper and explains why grasping from camera measurements requires this transform; GRASP-005 also converts depth-image grasps to world coordinates using camera parameters and known robot-camera calibration. | Need SAH-BRI-Grasp camera, target, station count, solver, and acceptance metrics. |
| Shared autonomy | Shared autonomy can reduce EEG control burden while the robot handles low-level planning and grasp execution. | inferred | README-001; BCI-004; MI-009; BRI-002; BRI-003; BRI-004; BRI-005; BRI-006; GRASP-004; GRASP-007; SA-001; SA-002 | BCI-004 supports modular online runtime/logging discipline; MI-009 verifies noninvasive high-level EEG commands fused with robot finite-state autonomy; BRI-002 verifies BCI manipulation benefits from vision, intent inference, and arbitration; BRI-003 supports behavior-tree task structuring for noisy teleoperation inputs; BRI-004 verifies a noninvasive AR BRI with gaze+MI high-level control and robot policy execution; BRI-005 and BRI-006 provide invasive direct-control upper bounds; GRASP-004 and GRASP-007 support robot-side grasp-ranking/servoing evidence and failure-mode logging; SA-001 and SA-002 verify shared-autonomy formulations and risks. Transfer to SSVEP-MI scene-aware grasping remains an inference. | Need SAH-BRI-Grasp baseline comparison, workload results, and failure-mode logs. |
| Shared-control formalism | Dynamic command-space control should be compared against shared-control formulations such as hindsight optimization or policy blending. | verified | SA-001; SA-002 | SA-001 provides a hindsight-optimization shared-autonomy baseline; SA-002 provides a policy-blending/arbitration baseline and shows wrong predictions can harm performance when assistance is too aggressive. | Need Exp3 implementation details for fixed command space vs scene-generated command space. |
| Dynamic command space | Dynamic command space is a central contribution rather than a simple integration detail. | inferred | README-001; YOLO-001; YOLO-002; YOLO-003; YOLO-004; YOLO-005; YOLO-007; YOLO-009; YOLO-010; BRI-002; BRI-003; BRI-004 | README positions YOLO outputs as command-space generation; YOLO-001/002/003/009/010 support visual candidate generation through YOLO-family detector evolution; YOLO-004, YOLO-005, and YOLO-007 add two-stage, natural-context dataset, and dense-detector comparator evidence; BRI-002, BRI-003, and BRI-004 support vision-assisted or structured shared-control BCI/BRI manipulation. Novelty remains a synthesis claim because prior comparators use different input and scene-binding mechanisms. | Need comparison against fixed command-space BCI and gaze/AR BRI systems. |
| Low-channel translation | 16/12/8/6/3 channel ablation can test wearable EEG translation potential. | inferred | README-001; BCI-005; SSVEP-005; SSVEP-010; SSVEP-016; SSVEP-020; SSVEP-021; SSVEP-022; SSVEP-023; SSVEP-026; SSVEP-031; MI-002; MI-004; MI-007; MI-008; MI-009; MI-014; MI-016; MI-021; MI-024; MI-025; MI-027; BRI-001; BRI-004 | SSVEP additions include few-electrode practical VEP/SSVEP systems, 8-electrode public datasets/models, and AR/consumer-headset evidence; MI additions include CSP/Riemannian/deep-learning baselines, 64-channel PhysioNet deep MI evidence, and wearable MI review evidence. Together they justify channel-ablation experiments, but do not verify the project's exact low-channel headset or montage. | Need local EEG hardware constraints and offline ablation results. |
| Rehabilitation product boundary | BCI-robotics rehabilitation is a possible product direction, but the current SAH-BRI-Grasp corpus does not yet justify clinical efficacy claims. | inferred | MI-030; MI-031; MI-032 | MI-030 supports user-tailored MI BCI feasibility in severely motor-impaired patients; MI-031 reviews BCI-robotics hand rehabilitation after stroke and notes many systems remain prototype/preclinical; MI-032 reports meta-analysis evidence for post-stroke upper-limb BCI training benefits. These support a future rehab direction, not current SAH-BRI-Grasp therapeutic claims. | Need separate clinical protocol, ethics, patient cohort, endpoints, and rehabilitation partner before any clinical product claim. |