独立行业技术评审 · 持续核验

TableVerse-100K

TableVerse-100K 是从非结构化互联网单帧重建的公开 Real2Sim 桌面操作数据:官方卡声明100K物理一致场景、近100万对象、超过35K语义类别,并附连续无碰撞专家 pick-and-place 轨迹。固定 HF 版本公开、非门控,数据卡声明 CC-BY-4.0;场景包含 MuJoCo XML/GLB、资产、纹理和重标注。它适合仿真操作/场景泛化研究,不是人形真机数据、可直接部署的 VLA 数据或实机迁移证明。

TableVerse-100K

TableVerse-100K 是从非结构化互联网单帧重建的公开 Real2Sim 桌面操作数据:官方卡声明100K物理一致场景、近100万对象、超过35K语义类别,并附连续无碰撞专家 pick-and-place 轨迹。固定 HF 版本公开、非门控,数据卡声明 CC-BY-4.0;场景包含 MuJoCo XML/GLB、资产、纹理和重标注。它适合仿真操作/场景泛化研究,不是人形真机数据、可直接部署的 VLA 数据或实机迁移证明。

开放状态
已核验开放
许可证
CC-BY-4.0 for the TableVerse Hugging Face dataset card declaration; repository code/documentation is separately Apache-2.0
商业使用
CC-BY-4.0 attribution obligations apply to the dataset declaration; separately audit source internet imagery, reconstructed assets, texture terms, target robot SDK and any unbundled model/API terms before commercial use
人工评审结论
official_public_nongated_cc_by_4_real2sim_synthetic_tabletop_dataset_code_pipeline_not_released_no_real_robot_transfer_verified
机器人
MuJoCo tabletop manipulation simulation; no physical robot platform disclosed
任务
instruction-annotated tabletop pick-and-place and relation-conditioned placement in dense, stacked and nested clutter
模态
single internet reference image provenance、3D GLB scene、MuJoCo XML、scene assets、textures、object/category relabel JSON、physical attributes、language task annotation、expert manipulation trajectory
格式
Hugging Face tar archives; documented per-scene files include assets/, textures/, scene.glb, scene.xml, scene_with_plane.xml and relabel_name.json. Full trajectory/action schema requires sampled-download validation
采集/生成方式
automated Real2Sim from in-the-wild single internet images: open-vocabulary detection, segmentation, mesh reconstruction, depth/6DoF alignment, layout-consistent collision rectification, MuJoCo settling, MLLM curation and cuRobo trajectory synthesis
已知问题
Code repository at reviewed fixed commit contains documentation/LICENSE but no executable Real2Sim, training, generation or evaluation implementation. Data card explicitly warns of long-tail geometry/semantic errors. Do not infer physical robot action semantics, full archive integrity, attribution chain, VLA readiness or real-robot success from the published card.

论文原文 · 官方项目页 · 官方代码固定提交 · 官方数据固定版本