# GRASP-004: Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics > 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-004.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [10.15607/RSS.2017.XIII.058](https://doi.org/10.15607/RSS.2017.XIII.058) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## GRASP-004: Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics ## Metadata * ID: GRASP-004 * Title: Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics * Year: 2017 * DOI / URL: 10.15607/RSS.2017.XIII.058 * 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: robust grasp planning from synthetic point clouds and analytic grasp metrics * Participants or dataset: Dex-Net 2.0 synthetic dataset with 6.7 million point clouds, grasps, and analytic metrics; physical robot trials on known and novel objects * Hardware: ABB YuMi robot, depth camera with known intrinsics, custom gripper setup * Channels or sensors: depth images / 2.5D point clouds and sampled grasp candidates ## Methods * Paradigm: sample antipodal grasp candidates from depth data and rank them by predicted robustness * Signal processing or model: Grasp Quality Convolutional Neural Network (GQ-CNN) trained on synthetic point clouds and robust grasp metrics * Training/calibration: synthetic training from thousands of 3D models; camera-robot calibration used in physical experiments * Online/offline: physical grasp-planning evaluations plus synthetic model validation ## Results * Metrics: grasp success rate, precision, planning time, physical trial performance * Main findings: GQ-CNN trained on Dex-Net 2.0 planned grasps in about 0.8 s, achieved 93% success on known adversarial objects, had the highest success on novel rigid objects, and achieved 99% precision on 40 novel household objects in reported tests * Reported limitations: failures were associated with missing depth data for thin geometry and collisions misclassified as robust; performance could improve with better depth sensing and analytic collision pruning ## Relevance To This Project * Supports: robust grasp-planning module design, depth/grasp-candidate ranking, and failure taxonomy for physical grasp execution * Conflicts with: does not include BCI intent, SSVEP/MI selection, or scene-aware command-space generation * Design implication: SAH-BRI-Grasp should keep grasp robustness, depth quality, collision checks, and calibration records as robot-side evidence distinct from EEG selection evidence ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Dex-Net 2.0 trains a GQ-CNN on synthetic point clouds, grasps, and analytic grasp metrics. | verified | The abstract describes a dataset of 6.7 million point clouds, grasps, and metrics used to train GQ-CNN. | Abstract | | GQ-CNN ranks grasp candidates from depth images for robust grasp planning. | verified | The architecture samples grasp candidates and uses GQ-CNN to determine the most robust grasp. | Introduction; Fig. 1 | | The method was evaluated in over 1,000 physical robot trials. | verified | The paper reports over 1,000 trials on an ABB YuMi comparing grasp-planning methods. | Abstract; Experiments | | Depth sensing and collision errors are important grasp failure modes. | verified | The discussion identifies missing sensor data for thin object geometry and collisions misclassified as robust as common failures. | Discussion / Future Work | | Dex-Net-like grasp ranking is a plausible robot-side autonomy comparator for SAH-BRI-Grasp. | inferred | The paper supports robust grasp planning, but not noninvasive EEG-driven scene-aware grasping. | Experiments; Discussion | ## Open Questions * Does project hardware provide depth quality sufficient for Dex-Net-like or GQ-CNN-like grasp ranking? * Should Exp3 log depth failures and collision-pruning failures as distinct robot-side failure modes? * Is the first prototype better served by a lighter geometric grasp method or a learned grasp-rank model?