Adversarial Differentiable Rendering for Grasp Quality Convolutional Neural Networks
Journal of Autonomous Vehicles and Systems · 01.10.2026 · 6 (4)
DOI: 10.1115/1.4072120 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Autonomous Vehicles and Systems, online erschienen am 23.09.2026. Verfasst von Hailie Mitchell und 3 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben.
Abstract (Originalsprache)
Adversarial example generation helps to determine the robustness of any given artificial neural network (ANN) dependent system, especially perception-dependent systems due to their large and complex input space, such as those with LiDAR or camera sensors. We develop a new attack based on differentiable rendering to find 3D adversarial inputs that mislead an artificial neural network commonly used in robotics, a grasp quality neural network. This neural network predicts the quality of hypothetical robotic parallel-jaw grasps. Our generated inputs are in the form of physically realizable 3D objects. Previous work either generated objects that were not necessarily feasible or relied on hand-crafted rules. We optimize the shape of existing 3D objects such that the model makes an incorrect prediction of the quality of a particular grasp on that object. Our attack is built on both a novel differentiable implementation of the physics-based Canny-Ferrari grasp quality and a novel self-collision avoidance strategy. With a thorough ablation study of grasps on 79 individual objects, we demonstrate that our attack generates the strongest adversarial objects (i.e., makes the model prediction furthest from ground truth).
Zitieren
APA
Mitchell, H., Tabor, T., Stein, M. V., & Goues, C. L. (2026). Adversarial Differentiable Rendering for Grasp Quality Convolutional Neural Networks. Journal of Autonomous Vehicles and Systems, 6(4). https://doi.org/10.1115/1.4072120
Vancouver
Mitchell H, Tabor T, Stein MV, Goues CL. Adversarial Differentiable Rendering for Grasp Quality Convolutional Neural Networks. Journal of Autonomous Vehicles and Systems. 2026;6(4). doi:10.1115/1.4072120
BibTeX
@article{Mitchell2026,
author = {Mitchell, Hailie and Tabor, Trenton and Stein, Meriel von and Goues, Claire Le},
title = {Adversarial Differentiable Rendering for Grasp Quality Convolutional Neural Networks},
journal = {Journal of Autonomous Vehicles and Systems},
year = {2026},
volume = {6},
number = {4},
doi = {10.1115/1.4072120}
}
EUVIMED