字节跳动 TableVerse:从互联网图片到 10 万个可仿真桌面操作场景 | Paper: 100K Simulation-Ready Tabletop Scenes

【中文】论文速递|字节跳动 TableVerse:从互联网图片到 10 万个可仿真桌面操作场景

机器人操作研究长期受限于大规模高保真场景数据:文本生成布局或程序化生成往往物理不合理、复杂度不足。字节跳动提出 TableVerse,一条 Real2Sim 自动化流水线:从无脚本的互联网图片重建真实桌面场景,生成物理一致、尺寸准确的仿真环境,并用任务条件化轨迹框架自动合成无碰撞的抓放演示,最终产出含 10 万个独特环境的 TableVerse-100K 数据集,覆盖多样的物体组合与真实空间分布。

【EN】Paper Brief | TableVerse: 100K Simulation-Ready Tabletop Scenes Reconstructed from Internet Images (ByteDance)

Robotic manipulation research is bottlenecked by large-scale, high-fidelity scene data — text-to-layout hallucination and procedural generation tend to be physically implausible and under-complex. ByteDance's TableVerse is an automated Real2Sim pipeline that reconstructs authentic tabletop scenes from unscripted internet images, producing physically consistent, accurately measured simulation environments, then synthesizes collision-free pick-and-place demonstrations via a task-conditioned trajectory framework — yielding the TableVerse-100K dataset of 100,000 unique environments.

来源 Source:

https://arxiv.org/abs/2607.21017

https://huggingface.co/papers/2607.21017