# SA-002: A policy-blending formalism for shared control

> 本页由本地论文卡片生成。请不要直接编辑本页；修改源卡片后运行 `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。

> 内部 wiki 边界：本页提供论文卡片证据和用户提供的 PDF 下载入口，仅用于私有阅读；全文抽取文本不发布。

> 来源: `library/paper_cards/SA-002.md`

## 论文访问

* 内部 PDF: <a href={"/papers/SA-002.pdf"} download style={{ display: "inline-flex", alignItems: "center", justifyContent: "center", minHeight: "2.25rem", padding: "0.45rem 0.8rem", borderRadius: "6px", backgroundColor: "#047857", color: "#ffffff", fontWeight: 700, lineHeight: 1, textDecoration: "none", boxShadow: "0 1px 2px rgba(15, 23, 42, 0.22)" }}>下载论文 PDF</a>
* DOI / 官方页面: [10.1177/0278364913490324](https://doi.org/10.1177/0278364913490324)
* 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。

## SA-002: A policy-blending formalism for shared control

## Metadata

* ID: SA-002
* Title: A policy-blending formalism for shared control
* Year: 2013
* DOI / URL: 10.1177/0278364913490324
* Local PDF: 见上方论文访问区块
* Text artifact: local-only path withheld from docs site
* Review status: `extracted`

## Study Type

* Track: Shared Autonomy / Shared Control
* Task: policy blending for shared-control teleoperation
* Participants or dataset: 8 participants; students; 4 male and 4 female; robot or video-game experience; no prior system experience
* Hardware: whole-body interface with OpenNI skeleton tracking and a robotic manipulator
* Channels or sensors: no EEG/BCI sensors

## Methods

* Paradigm: participants teleoperated an anthropomorphic robot to grasp tabletop objects; compared timid versus aggressive assistance under easy/hard and right/wrong prediction conditions
* Signal processing or model: user policy U and robot policy P blended by state-dependent arbitration alpha; goal prediction includes amnesic and memory-based variants; arbitration confidence may be distance-, probability-, entropy-, or probability-difference-based
* Training/calibration: task practice and user study; no BCI calibration
* Online/offline: online teleoperation user study plus post-experimental prediction analysis

## Results

* Metrics: hard tasks took 22.9 s longer than easy tasks, F(1,53)=18.45, p\<.001; wrong prediction took 30.1 s longer than right prediction, F(1,53)=31.88, p\<.001; aggressive assistance added 19.4 s overall, F(1,53)=13.2, p=.001, with interaction effects; memory-based prediction outperformed amnesic prediction, F(1,112)=1020.95, p\<.001
* Main findings: assistance effectiveness depends on prediction correctness, task difficulty, and arbitration aggressiveness; strong assistance is harmful when prediction is wrong
* Reported limitations: formalism covers continuous inputs modified by the robot; small laboratory user study; participants were not disaster responders or disabled users; results should be treated as trends rather than universal conclusions

## Relevance To This Project

* Supports: confidence-dependent arbitration and careful handling of wrong intent predictions
* Conflicts with: no EEG/BCI input and no discrete SSVEP/MI command stream
* Design implication: the SAH-BRI-Grasp state machine should use conservative arbitration until target confidence is high, and should expose cancel/recovery paths for wrong selections

## Extracted Evidence

| Claim | Status | Evidence Note | Page/Section |
| --- | --- | --- | --- |
| Shared-control arbitration should depend on confidence in the predicted goal. | verified | The formalism blends user and robot policies with state-dependent arbitration and discusses confidence measures. | The components of assistance |
| Incorrect intent prediction can make aggressive assistance worse. | verified | Wrong-prediction conditions strongly favored timid assistance and increased completion time. | A study on assistance; Analysis |
| Memory-based goal prediction improves over amnesic prediction in the reported analysis. | verified | The post-experimental analysis reports a significant effect favoring memory-based prediction. | Analysis |
| SAH-BRI-Grasp should keep robot assistance conservative until SSVEP/MI intent is sufficiently confirmed. | inferred | This applies the wrong-prediction/arbitration evidence to discrete EEG-driven object selection. | Limitations; Analysis |

## Open Questions

* What confidence signal should drive arbitration: SSVEP posterior, detector confidence, grasp feasibility, or a fusion of all three?
* How should the system recover from a confidently wrong target selection?
* How should user preference be balanced against faster task completion?