The University of California, Berkeley’s Laboratory for Automation Science and Engineering has released their Dexterity Network (Dex-Net) 2.0 dataset and accompanying neural network model to help researchers train robotic arms to grip a wide variety of objects. The dataset contains 6.7 million grasps, grasp robustness labels, and synthetic point clouds, which are sets of points on a 3D coordinate system. Researchers can use the data to train a neural network model to analyze the probability of a grasp successfully gripping, lifting, and carrying an irregular shaped object. Normally, grasping training datasets use fixed images and objects to train models and neural networks, but this dataset and the model improve upon this method by using the probabilistic success of a hypothetical action to grasp an object. This approach could help researchers develop AI and robotics systems that can better interact with unfamiliar objects.
Image: Oleg Alexandrov.
