# SA-001: Shared Autonomy via Hindsight Optimization

> Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`.

> Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published.

> Source: `library/paper_cards/SA-001.md`

## Paper Access

* Internal PDF: <a href={"/papers/SA-001.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)" }}>Download Paper</a>
* DOI / official page: [10.15607/RSS.2015.XI.032](https://doi.org/10.15607/RSS.2015.XI.032)
* Deployment boundary: these PDF links are intended only for a private/protected internal wiki.

## SA-001: Shared Autonomy via Hindsight Optimization

## Metadata

* ID: SA-001
* Title: Shared Autonomy via Hindsight Optimization
* Year: 2015
* DOI / URL: 10.15607/RSS.2015.XI.032
* Local PDF: see Paper Access section above
* Text artifact: local-only path withheld from docs site
* Review status: `extracted`

## Study Type

* Track: Shared Autonomy / Shared Control
* Task: shared autonomy for teleoperated robotic manipulation with uncertain user goals
* Participants or dataset: 10 user-study participants; 9 male and 1 female; robotics experience; no prior system experience
* Hardware: robot arm and Razer Hydra dual-joystick input
* Channels or sensors: no EEG/BCI sensors

## Methods

* Paradigm: users teleoperated a robot arm to grasp canteen/block/cup objects; compared policy method against predict-then-blend assistance
* Signal processing or model: goal-uncertain POMDP formulation; MaxEnt IOC estimates goal distribution from input history; QMDP/hindsight optimization approximates action selection; multi-goal/multi-target value functions support multiple grasp poses per object
* Training/calibration: user-study practice and method comparison; no BCI calibration
* Online/offline: online shared-autonomy user study

## Results

* Metrics: policy method was faster than blend, ANOVA F(1,9)=12.98, p=0.006; user input was lower, F(1,9)=7.76, p=0.021; preference evidence did not support the policy method over blend
* Main findings: assisting over a goal distribution can improve performance relative to committing to one predicted goal, but users may still prefer simpler or more controllable assistance
* Reported limitations: policy behavior may feel less controllable; user dissatisfaction and individualized costs need study; the user model assumes the user does not adapt input in response to assistance

## Relevance To This Project

* Supports: treating robot autonomy as goal-distribution assistance under uncertain intent
* Conflicts with: no EEG/BCI input; joystick teleoperation is higher bandwidth than SSVEP/MI commands
* Design implication: scene-aware command space should maintain uncertainty over candidate goals and use confirmation/recovery states rather than immediately committing to one object

## Extracted Evidence

| Claim | Status | Evidence Note | Page/Section |
| --- | --- | --- | --- |
| Shared autonomy can assist manipulation by reasoning over uncertain user goals. | verified | The paper formulates shared autonomy as a goal-uncertain POMDP and uses hindsight optimization to select assistive actions. | Problem Statement; Hindsight Optimization |
| Goal-distribution assistance can reduce execution time and user input. | verified | The user study reports faster execution and lower user input for the policy method than predict-then-blend. | User Study |
| Better task performance does not guarantee stronger user preference. | verified | Preference hypotheses were not supported; users tended to prefer or accept blend despite policy performance advantages. | User Study / Discussion |
| SAH-BRI-Grasp should include explicit user confirmation and controllability checks. | inferred | This follows from the performance/preference gap and the risk that assistance may feel less controllable. | Discussion |

## Open Questions

* How should SSVEP target probabilities and MI confirmations feed a shared-autonomy goal distribution?
* What confidence threshold should freeze a target candidate before robot motion?
* How should user preference and perceived control be measured in Exp3?