SAGE-10k
收藏资源简介:
SAGE-10k是由英伟达牵头,联合多所高校开发的仿真就绪3D场景数据集,专为具身智能体训练设计。该数据集包含10,000个通过代理框架生成的多样化室内场景,涵盖卧室、厨房等环境,每个场景均经过物理稳定性验证和语义合理性审核,支持开放词汇提示下的动态生成。数据通过多级增强技术(对象配置、类别和布局)实现规模化构建,并集成Isaac Sim物理引擎进行碰撞检测。其核心应用是解决机器人策略学习中真实数据获取成本高、安全性差的问题,为移动操作、抓取放置等任务提供可扩展的仿真训练环境。
SAGE-10k is a simulation-ready 3D scene dataset led by NVIDIA and co-developed in partnership with multiple universities, specifically tailored for embodied AI agent training. It contains 10,000 diverse indoor scenes generated via agent frameworks, covering typical indoor environments such as bedrooms and kitchens. Each scene has been validated for physical stability and reviewed for semantic rationality, and supports dynamic generation under open-vocabulary prompts. The dataset is scaled up through multi-level enhancement techniques including object configuration, category definition and layout arrangement, and integrates the Isaac Sim physics engine for collision detection. Its core application is to address the challenges of high acquisition cost and poor safety in real-world data collection for robot policy learning, providing scalable simulation training environments for tasks such as mobile manipulation, grasping and placement.
- 1SAGE: Scalable Agentic 3D Scene Generation for Embodied AI英伟达; 伊利诺伊大学厄巴纳-香槟分校; 康奈尔大学; 斯坦福大学 · 2026年



