Reliable grasp proposal generation for thin objects in EdgeGraspNet

Develop a reliable grasp-proposal strategy for EdgeGraspNet on thin objects that avoids dependence on directly modifying point-cloud normals or changing the method’s central assumption that grasps are parameterized by contact normals.

Background

The paper reports that EdgeGraspNet frequently fails to generate collision-free grasps for thin objects such as pliers and screwdrivers because its contact-normal-based sampling strategy depends on accurately recovering side surfaces and local surface normals from depth observations. The authors attempted to improve this behavior through viewpoint changes, point-cloud filtering, and proposal adjustment after generation, but these interventions did not produce reliable grasps.

The unresolved issue is therefore to obtain dependable grasp proposals for thin geometries while retaining EdgeGraspNet’s existing point-cloud representation and contact-normal parameterization. The paper’s own method addresses the failure by supplementing contact-based proposals with top-down grasps, but the cited passage leaves the corresponding improvement to EdgeGraspNet unresolved.

References

EdgeGraspNet exhibits a particularly clear proposal-generation failure on pliers and screwdrivers, where it achieves 0.0\% success and 0.0\% scene-clear rate. On these thin objects, it rarely proposes any collision-free grasps, and the few valid proposals fail. We attribute this failure mode to its contact-normal-based grasp-candidate sampling, a common strategy in current methods~\citep{icg_net,edgegraspnet,contact_graspnet}. This heuristic works well for bulky objects with clear side surfaces, but struggles on thin geometries, where side normals and sharp edges are difficult to recover from depth. Our model avoids this failure mode by complementing contact-based proposals with top-down grasps, which do not require estimated normals.~\footnote{We did not anticipate this result, as we initially included top-down grasp proposals only to increase data-collection diversity. We tried to improve the performance of EdgeGraspNet by changing viewpoint, filtering the pointcloud differently, and adjusting proposals after generation. However, we could not recover reliable grasps without either modifying the pointcloud normals directly or changing EdgeGraspNet's central assumption that grasps are parameterized by contact normals.}

— Continual Learning for 6-DoF Grasp Synthesis via Experience and Demonstrations  (2610.01301 - Schiavi et al., 1 Oct 2026) in Section 4, subsection “Real-World Experiments,” paragraph “Baseline Performance and Single-Category Adaptation,” footnote associated with the discussion of EdgeGraspNet’s failure on pliers and screwdrivers