ActiveViewPose-200K; ActiveManip-Bench
收藏资源简介:
ActiveViewPose-200K是由北京通用人工智能研究院构建的大规模语义相机控制数据集,包含20万条高质量图像-语言-相机运动配对数据,具有精细的任务导向标注。该数据集通过多模态对齐技术构建,支持机器人学习任务相关的视角调整能力。ActiveManip-Bench是首个针对主动操作的仿真基准,涵盖12类复杂任务、100个家居对象和20种场景配置,解决了固定视角评估的局限性。两者共同服务于机器人主动感知与操作系统的研发,推动视觉-语言-动作模型的场景适应能力突破。
ActiveViewPose-200K is a large-scale semantic camera control dataset constructed by the Beijing Academy of General Intelligence. It contains 200,000 high-quality paired data of image, language and camera motion, with fine-grained task-oriented annotations. Built via multimodal alignment technologies, this dataset supports the learning of task-related viewpoint adjustment capabilities for robots. ActiveManip-Bench is the first simulation benchmark for active manipulation, covering 12 types of complex tasks, 100 household objects and 20 scene configurations, which addresses the limitations of fixed-viewpoint evaluation. Together, the two datasets serve the research and development of robotic active perception and operating systems, and promote breakthroughs in the scene adaptation capabilities of vision-language-action models.




