RSS 2015 Workshop

Bridging the Gap between Data-driven and Analytical Physics-based Grasping and Manipulation

July 16, Rome

"Leveraging Big Data for Grasp Planning", Jeannette Bohg, Max Planck Institute for Intelligent Systems

In the data-driven methodology towards robotic manipulation, predictive models are learned from labeled training data. They typically infer the quality of a grasp candidate under real-world conditions such as incomplete knowledge of the object as well as uncertainty and noise in sensing and actuation. Thereby, these models go beyond classical stability metrics that have nice theoretical properties but have been shown to be unreliable predictors of grasp success under the aforementioned conditions. Different from the areas of computer vision or speech recognition, there are no large-scale, labeled databases for robotic grasping and manipulation. This however, seems to be the key to the state-of-the-art performance of methods that are grouped under the term deep learning. We propose a new large-scale database containing grasps that are applied to a large set of objects from numerous categories. These grasps are generated in simulation and are annotated with the standard epsilon and a new physics-metric. We use a descriptive and efficient representation of the local object shape at which the grasp is applied. Each grasp is annotated with the proposed metrics and representation. Given this data, we present a two-fold analysis: (i) We use crowdsourcing to analyze the correlation of the two metrics with grasp success as predicted by humans. The results confirm that the proposed physics metric is a more consistent predictor for grasp success than the epsilon metric. Furthermore it supports the hypothesis that human labels are not required for good ground truth grasp data. Instead the physics metric can be used for simulation data. (ii) We apply big-data learning techniques (Convolutional Neural Networks and Random Forests) to show how they can leverage the large-scale database for improved prediction of grasp success.

Exemplar-based Prediction of Object Properties from Local Shape Similarity