Robust Grasp Planning Over Uncertain Shape Completions

Abstract

We present a method for planning robust grasps over uncertain shape completed objects. For shape completion, a deep neural network is trained to take a partial view of the object as input and outputs the completed shape as a voxel grid.The key part of the network is dropout layers which are enabled not only during training but also at run-time to generate a set of shape samples representing the shape uncertainty through Monte Carlo sampling. Given the set of shape completed objects, we generate grasp candidates on the mean object shape but evaluate them based on their joint performance in terms of analytical grasp metrics on all the shape candidates. We experimentally validate and benchmark our method against another state-of-the-art method with a Barrett hand on 90000 grasps in simulation and 200 grasps on a real Franka Emika Panda. All experimental results show statistically significant improvements both in terms of grasp quality metrics and grasp success rate, demonstrating that planning shape-uncertaintyaware grasps brings significant advantages over solely planning on a single shape estimate, especially when dealing with complex or unknown objects.

Beyond Top-Grasps Through Scene Completion

“…To date, many grasping methods rely fully or in part on deep learning. Some methods only use deep learning to extract additional information about objects with e.g., shape completion [9], [10] or tactile information [11] and then use analytical methods to plan the actual grasp [12], while others employ data-driven grasp planning in an endto-end fashion to generate grasps directly from images [1]- [8]. We will review both shape completion and end-to-end data-driven grasp planning as both are vital parts of our grasping pipeline.…”

“…In the context of shape completion from incomplete pointclouds, most recent improvements come from the adoption of deep learning. For instance, different works have explored tailored network structures [9], [13], [14], semantic object classification to aid the reconstruction [15], the integration of other sensing modalities such as tactile information [11], or the exploitation of the network uncertainty [10].…”

“…In the context of robotics grasping, [9], [11] and [10] are the most interesting as they not only focus on shape reconstruction quality but also on grasping accuracy. In this work, we make use of our previous shape completion network [10] to complete objects but instead of planning grasps with analytical methods-which is computationally expensive-we turn to data-driven grasp planning.…”

“…Shape completion refers to reconstructing the shape of an object from partial information about it in the form of a point-cloud. More precisely, a shape completion algorithm estimates (O, p) given a point-cloud r. To shape complete objects we used the pre-trained fully convolutional hour-glass shaped Deep Neural Network (DNN) proposed in [10] whose input is a voxel grid of the point-cloud captured from the camera and output is a completed voxel grid. The completed voxel grid is post-processed into a mesh by merging it with the input point-cloud and running the marching cube algorithm [21].…”

“…The DNN in [10] also included dropout layers throughout that were active during run-time to generate a set of shape samples representing, through Monte Carlo sampling, the shape uncertainty. In this work, we also generate shape samples but average them together to get a mean shape, effectively ignoring the shape uncertainty.…”