Grasp-Anything
收藏资源简介:
Grasp-Anything是一个由基础模型合成的大规模抓取数据集,旨在解决机器人抓取检测中的长期挑战。该数据集包含100万样本和超过300万个对象,通过利用基础模型的广泛知识库,涵盖了日常生活中的各种对象,从而超越了以往的数据集。Grasp-Anything不仅在多样性和规模上表现出色,还成功支持了基于视觉任务的零样本抓取检测和真实世界机器人实验。数据集的创建过程涉及使用ChatGPT生成场景描述,然后利用基础模型生成图像和抓取姿态。该数据集的应用领域包括机器人抓取、零样本学习和领域适应,旨在通过提供更广泛的对象和更自然的场景设置来提高抓取检测的泛化能力。
Grasp-Anything is a large-scale grasping dataset synthesized by foundation models, aiming to address long-standing challenges in robotic grasping detection. This dataset contains 1 million samples and over 3 million objects, covering a wide range of daily objects by leveraging the extensive knowledge base of foundation models, thus outperforming previous datasets. Grasp-Anything excels not only in diversity and scale, but also successfully supports zero-shot grasping detection for vision-based tasks and real-world robotic experiments. The dataset creation process involves using ChatGPT to generate scene descriptions, followed by generating images and grasping poses via foundation models. Its application areas include robotic grasping, zero-shot learning, and domain adaptation, with the goal of enhancing the generalization performance of grasping detection by providing a broader set of objects and more natural scene configurations.

- 1Grasp-Anything: Large-scale Grasp Dataset from Foundation Models自动化与控制研究所,维也纳工业大学,维也纳,奥地利 · 2023年



