Cloud-Based Grasp Analysis and Planning for Toleranced Parts Using Parallelized Monte Carlo Sampling

This paper considers grasp planning in the presence of shape uncertainty and explores how cloud computing can facilitate parallel Monte Carlo sampling of combination actions and shape perturbations to estimate a lower bound on the probability of achieving force closure. We focus on parallel-jaw push grasping for the class of parts that can be modeled as extruded 2-D polygons with statistical tolerancing. We describe an extension to model part slip and experimental results with an adaptive sampling algorithm that can reduce sample size by 90%. We show how the algorithm can also bound part tolerance for a given grasp quality level and report a sensitivity analysis on algorithm parameters. We test a cloud-based implementation with varying numbers of nodes, obtaining a 515 speedup with 500 nodes in one case, suggesting the algorithm can scale linearly when all nodes are reliable. Code and data are available at: http://automation.berkeley.edu/cloud-based-grasping. Note to Practitioners-In manufacturing, small variations in part shape are inevitable. This paper addresses the challenge of grasping parts with a parallel-jaw gripper where the true part shape is modeled with statistical tolerancing. We use sampling to compute the grasp position and orientation that optimizes the probability of achieving a stable grasp. The computation requires many samples but can be parallelized and performed efficiently using Cloud computing. We present algorithms and experiments using PiCloud, a commercial cloud computing platform. In future work, we hope to extend the 2-D linear analysis to 3-D parts with curved surfaces.