# GRASP-005: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach > 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/GRASP-005.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.15607/RSS.2018.XIV.021](https://doi.org/10.15607/RSS.2018.XIV.021) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## GRASP-005: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach ## Metadata * ID: GRASP-005 * Title: Closing the Loop for Robotic Grasping: A Real-time, Generative Grasp Synthesis Approach * Year: 2018 * DOI / URL: 10.15607/RSS.2018.XIV.021 * Local PDF: see Paper Access section above * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: Robot Grasping * Task: real-time closed-loop grasp synthesis for unknown and dynamic objects * Participants or dataset: Cornell Grasping Dataset for training; over 2000 real robot grasp attempts on adversarial, household, isolated, and cluttered object settings * Hardware: Kinova Mico 6DOF robot, Kinova KG-2 two-finger gripper, wrist-mounted Intel RealSense SR300 RGB-D camera, GPU desktop * Channels or sensors: depth image input for GG-CNN; robot pose and camera calibration are used for grasp execution ## Methods * Paradigm: predict antipodal grasp quality, angle, and gripper width at every depth-image pixel * Signal processing or model: Generative Grasping CNN maps an inpainted depth image to a grasp map without sampling individual grasp candidates * Training/calibration: trained from Cornell grasp annotations augmented with crops, zooms, and rotations; execution converts image-space grasps to world coordinates using camera intrinsics and known robot-camera calibration * Online/offline: real robot experiments include open-loop and closed-loop grasping; closed-loop control uses PBVS with depth images processed at 30 Hz ## Results * Metrics: grasp success rate, pipeline runtime, dynamic/clutter robustness * Main findings: full grasping pipeline runs in 19 ms on the reported GPU desktop; GG-CNN supports closed-loop grasp generation up to 50 Hz; dynamic-object success rates are 83% on adversarial objects and 88% on household objects; dynamic clutter success is 81% * Reported limitations: RealSense depth fails at close range and on some black/reflective objects; gripper geometry limits thin/large/heavy objects; execution still depends on camera parameters and known transforms ## Relevance To This Project * Supports: closed-loop visual grasping as the low-level autonomy layer after a high-level BCI object/action selection * Conflicts with: does not include BCI, SSVEP, MI, user intent, or semantic object selection * Design implication: SAH-BRI-Grasp should keep YOLO/BCI target selection separate from RGB-D grasp-pose generation and should log calibration, depth failures, and closed-loop recovery behavior ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | GG-CNN predicts grasp quality, angle, and width for every pixel in a depth image. | verified | The abstract, Fig. 1 description, and grasp-map definition specify pixelwise outputs Q, Phi, and W. | Abstract; Sections III-IV | | The system is designed for real-time closed-loop grasping. | verified | The abstract reports closed-loop control up to 50 Hz; the implementation section reports 19 ms for the full pipeline and 30 Hz depth-image processing in closed-loop PBVS. | Abstract; Sections V.A, V.D | | Dynamic-object and dynamic-clutter grasping are evaluated on a real robot. | verified | Experiments report dynamic grasping success on adversarial and household objects and 81% success in dynamic clutter. | Sections VI.B-VI.C | | The method still relies on camera intrinsics and known robot-camera calibration for world-coordinate execution. | verified | The grasp definition maps image-space grasps to world coordinates through camera parameters and known calibration. | Section III | | Closed-loop grasp synthesis is a plausible low-level autonomy module for SAH-BRI-Grasp. | inferred | The paper shows robot-side closed-loop grasp execution, but it does not integrate noninvasive EEG or scene-aware command generation. | Sections V-VII | | This card does not resolve the project's hand-eye calibration protocol. | needs confirmation | It states the need for known calibration but does not provide the project-specific calibration procedure or validation. | Sections III, V | ## Open Questions * Which available RGB-D camera and gripper constraints match the SAH-BRI-Grasp hardware? * Can project experiments reproduce closed-loop target updating while an SSVEP target is being presented? * What fallback is needed when depth is missing, target boxes shift, or grasp quality is low?