House3D Environment

House3D 是一个虚拟 3D 环境，由数千个室内场景组成，这些场景配备了来自 SUNCG 数据集的各种场景类型、布局和对象。它由超过 45,000 个室内 3D 场景组成，从工作室到带游泳池和健身房的两层房屋。所有 3D 对象都使用类别标签进行了完整注释。环境中的代理可以访问多种模式的观察结果，包括 RGB 图像、深度、分割掩码和自上而下的 2D 地图视图。渲染器以每秒数千帧的速度运行，使其适用于大规模 RL 训练。

House3D is a virtual 3D environment consisting of thousands of indoor scenes equipped with diverse scene types, layouts and objects sourced from the SUNCG dataset. It comprises over 45,000 indoor 3D scenes ranging from studio apartments to two-story houses with swimming pools and gyms. All 3D objects are fully annotated with category labels. Agents within this environment can access multimodal observations, including RGB images, depth maps, segmentation masks and top-down 2D map views. The renderer operates at thousands of frames per second, making it suitable for large-scale reinforcement learning (RL) training.

搜集汇总

背景与挑战

背景概述

House3D Environment是一个由45,000多个室内3D场景组成的虚拟环境，涵盖多种场景类型和布局，所有对象都有类别标签注释。它支持RGB图像、深度、分割掩码等多种观察模式，并以高速渲染能力适用于大规模强化学习训练。