Measuring Uncertainty in Shape Completion to Improve Grasp Quality

TL;DR

This work tackles uncertainty in single-view 3D shape completion for robotic grasping by estimating completion uncertainty with Monte Carlo Dropout on the 3DSGrasp network and integrating it into grasp scoring. The completed point cloud and per-point uncertainty are used to adjust grasp scores via $S' = S - W_u \sum_{p \in CP'} \sigma_u$, enabling re-ranked top-5 grasps. Empirical validation on a Kinova Gen3 with a 2F-85 gripper across 10 YCB objects shows higher rank-5 success rates than baselines (GPD, 3DSGrasp) and ablations, with reasonable computation time (~4s for completion, ~2s ranking). The results demonstrate that accounting for shape-prediction noise improves grasp reliability in occluded, real-world settings and motivates future uncertainty-aware grasp generators.

Abstract

Shape completion networks have been used recently in real-world robotic experiments to complete the missing/hidden information in environments where objects are only observed in one or few instances where self-occlusions are bound to occur. Nowadays, most approaches rely on deep neural networks that handle rich 3D point cloud data that lead to more precise and realistic object geometries. However, these models still suffer from inaccuracies due to its nondeterministic/stochastic inferences which could lead to poor performance in grasping scenarios where these errors compound to unsuccessful grasps. We present an approach to calculate the uncertainty of a 3D shape completion model during inference of single view point clouds of an object on a table top. In addition, we propose an update to grasp pose algorithms quality score by introducing the uncertainty of the completed point cloud present in the grasp candidates. To test our full pipeline we perform real world grasping with a 7dof robotic arm with a 2 finger gripper on a large set of household objects and compare against previous approaches that do not measure uncertainty. Our approach ranks the grasp quality better, leading to higher grasp success rate for the rank 5 grasp candidates compared to state of the art.

Figures (5)

• Figure 1: The proposed approach for measuring uncertainty in the shape completion models and including uncertainty in grasp generator algorithms. Our model receives the partial PCD and generates a completed PCD for the object shape and the PCD's std corresponding low/high std as low/high uncertainty points, respectively. The default grasp algorithm generates grasp candidates and rank it using a grasp score metric. We proposed an extension to include the uncertainty in the grasp score metric. *The same grasp candidates have a new ranking according to their new grasp quality score.
• Figure 2: Shape completion and respective uncertainty values for different partial PCD of objects present in the YCB dataset. These partials are segments including only surface points of the YCB object dataset. Top to bottom, left to right: french yellow mustard, mushroom, pringles original, black and decker lithium drill box, box, mini soccer ball, tomato soup, detergent bottle, donut.
• Figure 3: Adding the uncertainty in the grasp quality score of one grasp candidate (best ranked grasp) for the complete PCD of the pringles can. The algorithm iterates from left to right: (i) output of the shape completion network with uncertainty layer; (ii) output of GPD pas_grasp_2017 for the complete PCD; (iii) for each grasp, start with the first; (iv) find the 3D points in the complete PCD that are inside the gripper; (v) crop the 3D points and get the standard deviation of the points in the cropped PCD; (vi) add the uncertainty of those points to the grasp quality score.
• Figure 4: An example of our proposed grasp quality score showing the before and after output for the completed PCD of the pringles can. Before is the GPD output and after is our proposed algorithm. The after ranking index is [4, 1, 5, 2, 3] for the top 5 grasp pose candidates. Colors from best to worst: bright green, light green, yellow, orange, red.
• Figure 5: The rank 5 grasp candidates for each of the three methods: GPD, 3DSGrasp, and ours. The GPD method often leads to collisions and/or grasps that are unreachable by the robot; 3DSGrasp solves the issue with object collisions but still has many unreachable grasps; our method keeps the benefits of 3DSGrasp but produces less unreachable grasps.