InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts

The advancement of Embodied AI heavily relies on large-scale, simulatable 3Dscene datasets characterized by scene diversity and realistic layouts. However,existing datasets typically suffer from limitations in data scale or diversity,sanitized layouts lacking small items, and severe object collisions. To addressthese shortcomings, we introduce InternScenes, a novel large-scalesimulatable indoor scene dataset comprising approximately 40,000 diverse scenesby integrating three disparate scene sources, real-world scans, procedurallygenerated scenes, and designer-created scenes, including 1.96M 3D objects andcovering 15 common scene types and 288 object classes. We particularly preservemassive small items in the scenes, resulting in realistic and complex layoutswith an average of 41.5 objects per region. Our comprehensive data processingpipeline ensures simulatability by creating real-to-sim replicas for real-worldscans, enhances interactivity by incorporating interactive objects into thesescenes, and resolves object collisions by physical simulations. We demonstratethe value of InternScenes with two benchmark applications: scene layoutgeneration and point-goal navigation. Both show the new challenges posed by thecomplex and realistic layouts. More importantly, InternScenes paves the way forscaling up the model training for both tasks, making the generation andnavigation in such complex scenes possible. We commit to open-sourcing thedata, models, and benchmarks to benefit the whole community.