A Two-stage Learning Architecture that Generates High-Quality Grasps for a Multi-Fingered Hand

Abstract

We investigate the problem of planning stable grasps for object manipulations using an 18-DOF robotic hand with four fingers. The main challenge here is the highdimensional search space, and we address this problem using a novel two-stage learning process. In the first stage, we train an autoregressive network called the hand-pose-generator, which learns to generate a distribution of valid 6D poses of the palm for a given volumetric object representation. In the second stage, we employ a network that regresses 12D finger joint configurations and a scalar grasp quality from given object representations and palm poses. To train our networks, we use synthetic training data generated by a novel grasp planning algorithm, which also proceeds stage-wise: first the palm pose, then the finger positions. Here, we devise a Bayesian Optimization scheme for the palm pose and a physics-based grasp pose metric to rate stable grasps. In experiments on the YCB benchmark data set, we show a grasp success rate of over 83%, as well as qualitative results grasping unknown objects on a real robot system.

Learning-Based Real-Time Torque Prediction for Grasping Unknown Objects with a Multi-Fingered Hand

“…12 000 objects from the ShapeNet dataset [19]. For each object, we generate grasps using our grasping network [20]. Specifically, we let the network predict 1024 grasps per object.…”

“…We evaluate our approach on our precisely calibrated [22] humanoid robot Agile Justin [23] by incorporating it into our grasping pipeline: For a given object, we first observe an incomplete 3D model using a Kinect depth camera [24]. Via shape completion, the observation is completed and used to predict a stable grasp via our grasping network [20]. Using a learning-based motion planner [25], the hand now approaches the object as specified by the predicted grasp.…”

Shape Completion with Prediction of Uncertain Regions

“…To measure the effect of uncertain region prediction on object grasping, we use the network from [33] to predict 1024 grasps for a humanoid four-finger hand (DLR-Hand II [34]) per object from its completion. or the uncertaintyaware variant we use the predicted uncertain region as a filter, removing grasps that intersect with it.…”

“…or the uncertaintyaware variant we use the predicted uncertain region as a filter, removing grasps that intersect with it. We evaluate the grasps' quality using the Improved Epsilon Quality (IEQ) [33] metric, which measures the minimal external force applied to the object (ground truth mesh) that would break the grasp where the highest IEQ in the sample is the achieved grasp quality.…”