# GRASP-001: Data-Driven Grasp Synthesis - A Survey > 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/GRASP-001.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [10.1109/TRO.2013.2289018](https://doi.org/10.1109/TRO.2013.2289018) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## GRASP-001: Data-Driven Grasp Synthesis - A Survey ## Metadata * ID: GRASP-001 * Title: Data-Driven Grasp Synthesis - A Survey * Year: 2014 * DOI / URL: 10.1109/TRO.2013.2289018 * 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: Robot Vision / Grasping / Calibration * Task: survey of data-driven grasp synthesis * Participants or dataset: review paper; references existing grasp databases such as Columbia grasp database, VisGraB, and playpen, but no new unified dataset * Hardware: surveyed systems include 2D/3D vision, tactile sensors, PR2 sensor head, Kinect, stereo camera, point clouds, force-torque, and haptic feedback * Channels or sensors: varied across surveyed systems ## Methods * Paradigm: data-driven grasp synthesis by sampling grasp candidates and ranking them by metrics or learned experience * Signal processing or model: grasp candidates can be parameterized by grasping point, approach vector, wrist orientation, and initial finger configuration * Training/calibration: varies across reviewed approaches; known-object methods often build an offline object-grasp database and retrieve/filter grasps online * Online/offline: survey covers offline databases, online recognition/pose estimation, and execution-stage feedback ## Results * Metrics: no single experimental metric; this is a survey * Main findings: object detection or pose estimation is only part of grasping; a system still needs grasp-hypothesis generation, ranking, reachability filtering, and robust execution * Reported limitations: object segmentation is often simplified; task-dependent grasping and hand kinematics remain underdeveloped; real-world grasp success is hard to predict from classical metrics; simulation-to-real gaps and lack of comprehensive benchmarks remain open ## Relevance To This Project * Supports: separating object-candidate generation from grasp synthesis and execution * Conflicts with: no BCI or SSVEP/MI component * Design implication: after a BCI selects an object, the robot must still estimate pose/affordance, generate candidate grasps, filter by reachability and safety, and use feedback or confirmation ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Data-driven grasping involves sampling and ranking grasp candidates, not just detecting object boxes. | verified | The survey defines grasp synthesis as candidate generation plus ranking and discusses grasp parameterization. | Abstract; data-driven definition; grasp parameterization | | Object prior knowledge changes the grasping pipeline. | verified | The survey distinguishes known, familiar, and unknown object cases and describes known-object database/retrieval flows. | Known/familiar/unknown taxonomy | | Object detection should be separated from grasp-pose planning in SAH-BRI-Grasp. | inferred | This applies the survey's grasping pipeline boundary to YOLO-based candidate selection. | Known-object flow; open problems | | Current grasping evidence in this card does not verify hand-eye calibration. | needs confirmation | The survey discusses sensing and pose uncertainty but does not provide a specific calibration protocol for this project. | Open problems | ## Open Questions * Which first prototype grasp method is compatible with selected object IDs and available sensors? * Should the system use grasp rectangles, RGB-D point clouds, analytic grasp metrics, or a learned grasp network? * What minimum feedback is needed to recover from failed or unsafe grasp planning?