Embodied-Points-200K
收藏资源简介:
Embodied-Points-200K数据集是一个包含约20万个样本的高质量语料库,旨在提升机器人在机器人操作中的空间推理和指代能力。该数据集由天津大学的研究团队创建,旨在解决机器人操作中视觉感知与物理动作之间的“看见-做”差距问题。数据集内容来源于多种机器人操作数据集,支持四个关键指代能力:指代表达式定位、区域指代表达式定位、对象功能定位和视觉轨迹生成。数据集的创建过程使用了强化学习技术,并通过多任务奖励设计进行训练,从而实现了在多个空间理解和精确指代基准测试中的最先进性能。Embodied-Points-200K数据集的应用领域包括机器人操作、强化学习和视觉语言模型的研究。
The Embodied-Points-200K dataset is a high-quality corpus containing approximately 200,000 samples, designed to enhance robots' spatial reasoning and referring capabilities in robotic manipulation. This dataset was created by a research team from Tianjin University, aiming to address the "see-do" gap between visual perception and physical actions in robotic manipulation. The dataset content is derived from multiple robotic manipulation datasets, and it supports four core referring capabilities: referring expression grounding, regional referring expression grounding, object functionality grounding, and visual trajectory generation. The dataset was developed using reinforcement learning techniques, trained with a multi-task reward design, and achieved state-of-the-art performance across multiple spatial understanding and precise referring benchmarks. The application domains of the Embodied-Points-200K dataset include research in robotic manipulation, reinforcement learning, and vision-language models.




