DefGraspSim: Simulation-based grasping of 3D deformable objects

From clothing, to plastic bottles, to humans, deformable objects are omnipresent in our world. A large subset of these are 3D deformable objects (e.g., fruits, internal organs, and flexible containers), for which dimensions along all 3 spatial axes are of similar magnitude, and significant deformations can occur along any of them [46]. Robotic grasping of 3D deformables is underexplored relative to rope and cloth, but is critical for applications like food handling [14], robotic surgery [47], and domestic tasks [46]. Compared to rigid objects, grasping 3D deformable objects faces 4 major challenges. First, classical analytical metrics for grasping rigid objects (e.g., force/form closure) do not typically consider deformation of the object during or after the grasp [46]. Yet, deformations significantly impact the contact surface and object dynamics. Second, existing grasp strategies for rigid objects may not directly transfer to 3D deformables, as compliance can augment or reduce the set of feasible grasps. For example, we may grasp a soft toy haphazardly; however, if the toy were rigid, it would no longer conform to our hands, and many grasps may become unstable. Conversely, we may grasp a rigid container haphazardly; however, if the container were flexible, grasps along its faces may crush its contents. Third, the definition of a successful grasp on a 3D deformable is highly dependent on object properties, such as fragility and compliance. Thus, grasp outcomes must be quantified by diverse performance metrics, such as stress, deformation, and stability. Performance metrics may also compete (e.g., a stable grasp may induce high deformation). Fourth, performance metrics may be partially or fully unobservable (e.g., volumetric stress fields), requiring estimation in the real world. Previous works have typically formulated quality metrics, which we refer to more generally as grasp features: simple quantities that a robot can measure before pickup that can predict performance metrics. Whereas many grasp features have been proposed for rigid objects, analogous features for deformable objects are limited. Given these complexities, we conduct a large-scale simulation-based study of 3D deformable object grasping (Fig. 1). Simulation affords multiple advantages: it extends analytical methods through accurate modeling of object deformation, enables safe execution of experiments, and provides full observability of performance metrics. For an overview of existing literature on deformable object modeling, grasp performance metrics, and grasp features, see Appendix A.