# SAH-BRI-Grasp

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

> Source: `system/overview.md`

## SAH-BRI-Grasp System Overview

Status: `current system orientation; performance remains unverified`

`SAH-BRI-Grasp` is the current system instance of the broader `SAH-BRI` framework. It studies scene-aware hybrid SSVEP-MI brain-robot interaction for vision-guided shared-control robotic grasping.

## Core Research Question

Can a scene-aware hybrid BCI reduce the control burden of non-invasive EEG by moving low-level robot motion into shared autonomy, while keeping the user in the loop through high-level target selection, confirmation, intervention, and stopping commands?

The current answer is a system design and experiment plan, not a verified performance result.

## Research Transition

```text
from brain-controlled robot-arm commands
to brain-intent-guided robotic grasping
```

The project does not ask EEG to control every robot joint or continuous 6-DoF motion. It assigns low-bandwidth human intent and high-dimensional robot execution to different layers.

Claim status: `inferred` from the current component evidence. Integrated performance remains `needs confirmation`.

## System Flow

```text
RGB or RGB-D scene
  -> YOLO-style object detection
  -> candidate tracking and freeze
  -> object-bound SSVEP stimuli
  -> SSVEP target selection
  -> MI or SSVEP confirmation, intervention, pause, or stop
  -> confidence and safety arbitration
  -> grasp-pose generation and reachability checks
  -> robot approach and grasp execution
  -> visual or end-effector verification
  -> success, recovery, or safe abort
```

## Responsibility Split

| Layer | Responsibility | Current Status |
| --- | --- | --- |
| SSVEP | discrete scene-object selection and confirmation | `inferred`; dynamic object targets require Exp1 |
| Motor imagery | mode control, intervention, pause, stop, or confirmation | `inferred`; no-control safety requires Exp2 |
| Vision | detect scene objects and generate selectable candidates | detector role supported; project binding requires Exp1/Exp3 |
| Grasping and calibration | convert selected objects into executable grasp poses | project implementation and calibration remain `needs confirmation` |
| Shared autonomy | combine uncertain user intent with robot feasibility | `inferred`; baseline comparison requires Exp3 |
| Robot | perform low-level planning, motion, grasping, and recovery | hardware and middleware binding remain `needs confirmation` |

## Dynamic Command Space

Traditional BCI robot interfaces often expose a fixed command set:

```text
left | right | forward | backward | grasp | release
```

`SAH-BRI-Grasp` instead proposes that the physical scene generates the selectable command space:

```text
detected object
  -> stable candidate ID
  -> SSVEP stimulus assignment
  -> decoded semantic target
  -> robot-side grasp plan
```

Treating detector outputs as BCI commands is a project-level synthesis claim. It is not yet verified by the current local experiments.

## Decision And Safety Loop

No single decoder output should trigger unrestricted physical execution. The decision layer must consider:

* SSVEP and MI confidence;
* no-control and rejection states;
* target stability and target loss;
* grasp feasibility and reachability;
* calibration validity;
* collision and workspace constraints;
* pause, cancel, stop, and emergency-stop paths.

The detailed transition rules live in the [State Machine](/framework/state-machine), [Safety Design](/framework/safety-design), and [Runtime Interfaces](/framework/runtime-interfaces).

## Experiment Chain

| Experiment | Question | Claim Gate |
| --- | --- | --- |
| [Exp1 SSVEP-YOLO](/experiments/exp01-ssvep-yolo-target-selection) | Can scene objects become stable SSVEP-selectable targets? | dynamic scene target selection |
| [Exp2 MI Mode Control](/experiments/exp02-mi-mode-control) | Can MI support intervention without unsafe false activation? | mode-control and no-control safety |
| [Exp3 Closed-Loop Grasping](/experiments/exp03-closed-loop-grasping) | Can the full shared-autonomy loop complete physical grasping? | system task and grasp performance |
| [Exp4 Channel Ablation](/experiments/exp04-channel-ablation) | Which EEG channels are needed for reduced-channel translation? | low-channel feasibility |

All four experiments remain pre-data protocols.

## Current Evidence Boundary

The local corpus supports component-level rationale for:

* non-invasive EEG as a constrained intent channel;
* SSVEP target selection;
* MI decoding and mode-level robot control;
* real-time object detection;
* grasp synthesis and hand-eye calibration;
* hybrid BRI and shared autonomy.

The repository does not yet verify:

* project-specific task or grasp success;
* workload reduction;
* safe online MI false-activation rates;
* dynamic object-bound SSVEP performance;
* low-channel online performance;
* hardware-specific timing or calibration accuracy.

Use the [Claim Map](/research-map/claim-map), [Evidence Matrix](/evidence/evidence-matrix), and [Open Questions](/research-map/open-questions) before promoting any of these into manuscript claims.

## Current And Future Scope

The current system remains `SAH-BRI-Grasp`. A dexterous hand can be evaluated as an interchangeable end effector without changing the theme.

Only when task intention, skill selection, contact adaptation, or dexterous manipulation becomes a primary experimental variable should the project consider the provisional `SAH-BRI-Manip` system direction.

See [Dexterous Arm-Hand Extension](/research-map/dexterous-arm-hand-extension) for that boundary.

## Continue Reading

1. Review the [Research Map](/research-map) to connect the system to its claims.
2. Check [Open Questions](/research-map/open-questions) to see what is unresolved.
3. Use the [Experiment Roadmap](/experiments/roadmap) to follow the validation order.