# GRASP-003: Real-Time Grasp Detection Using Convolutional Neural Networks > 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-003.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.1109/ICRA.2015.7139361](https://doi.org/10.1109/ICRA.2015.7139361) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## GRASP-003: Real-Time Grasp Detection Using Convolutional Neural Networks ## Metadata * ID: GRASP-003 * Title: Real-Time Grasp Detection Using Convolutional Neural Networks * Year: 2015 * DOI / URL: 10.1109/ICRA.2015.7139361 * 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: RGB-D grasp rectangle detection with convolutional neural networks * Participants or dataset: Cornell Grasping Dataset with RGB-D object images and labeled grasps * Hardware: GPU evaluation; robot execution is not the main experimental focus * Channels or sensors: RGB-D images; grasp represented as oriented rectangles ## Methods * Paradigm: predict a five-dimensional grasp rectangle from an RGB-D image * Signal processing or model: single-stage CNN regression to grasp coordinates; regression plus classification; locally constrained MultiGrasp prediction for multiple grasp candidates * Training/calibration: trained on Cornell grasp labels with ImageNet pretraining; image-wise and object-wise evaluation splits * Online/offline: offline dataset evaluation of real-time grasp detection speed and accuracy ## Results * Metrics: rectangle grasp-detection accuracy, processing rate, object classification accuracy * Main findings: direct regression achieved about 85% accuracy; MultiGrasp achieved 88% grasp detection accuracy at 13 frames per second; combined regression/classification predicted seen object categories 90% of the time and novel object categories 61.5% of the time * Reported limitations: multiple-grasp prediction lacked a suitable quantitative dataset metric; the method assumes RGB-D grasp detection and does not validate full robot execution, calibration, or BCI integration ## Relevance To This Project * Supports: the boundary between object detection and grasp-rectangle prediction, and the need for a separate grasp-pose module after YOLO candidate selection * Conflicts with: does not use EEG, SSVEP, MI, shared autonomy, or physical hand-eye validation * Design implication: SAH-BRI-Grasp should not execute YOLO boxes directly; candidate objects need a grasp-pose estimator and robot feasibility checks ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Grasp detection can be formulated as predicting an oriented grasp rectangle from RGB-D input. | verified | The paper defines a five-dimensional grasp representation with position, orientation, height, and width. | Problem Description | | CNN-based grasp detection can run in real time on the reported setup. | verified | The abstract and results report 13 frames per second. | Abstract; Results | | MultiGrasp improved Cornell grasp detection performance in this paper. | verified | Results report 88% grasp detection accuracy and state-of-the-art performance at the time. | Results; Conclusion | | Grasp detection and object classification can be combined in one CNN pipeline. | verified | The regression plus classification model predicted object categories while maintaining grasp detection accuracy. | Results | | SAH-BRI-Grasp needs a grasp-pose module separate from YOLO object detection. | inferred | This follows from the paper's grasp-rectangle prediction task, but the paper does not evaluate SAH-BRI-Grasp. | Problem Description; Results | ## Open Questions * Will the project use geometric, GG-CNN-like, Dex-Net-like, or another grasp-pose method? * What RGB-D camera and calibration quality are required for executing predicted grasp rectangles? * How should failed grasp-pose predictions be logged separately from BCI selection errors?