KineBench Embodied World Model Benchmark
KineBench把生成视频经YOLO、MoGeV2与FoundationPose还原为6D末端轨迹，并在ManiSkill3闭环执行，覆盖20项操作任务和4套评测。官方MIT仓与CC-BY-2.0数据集真实公开、非gated；但README仅一行、无requirements，配置硬编码作者路径，必要权重、CAD和第三方仓未形成可复现闭包。
- 开放状态
- 数据可下载·上游条款待逐项确认
- 许可证
- Dataset CC-BY-2.0; repository code MIT; upstream simulator, models, checkpoints and assets retain separate licenses
- 商业使用
- dataset and code licenses permit broad reuse, but all upstream dependencies and generated-model service terms require separate review
- 人工评审结论
- verified_public_code_and_dataset_with_incomplete_environment_weights_and_assets
- 机器人
- ManiSkill3 simulated manipulation robots
- 任务
- 20 manipulation tasks across basic execution, task transfer, visual out-of-distribution generalization and complexity scaling
- 模态
- generated and reference videos、task annotations、language instructions、text embeddings、simulator trajectory references、6D end-effector evaluation outputs
- 格式
- Hugging Face repository、JSON annotations、video files、NumPy-compatible arrays、Python evaluation code
- 采集/生成方式
- ManiSkill3
- 已知问题
- README and dataset card are effectively empty; no requirements/environment lock; configs and third-party preparation script contain author-local paths; required YOLO/MoGe weights, CAD and third-party checkouts are absent; no physical validation or independent reproduction.