code_weights_and_dataset_major_update · 已核验官方资源
审核码：verified_official; 官方关系：verified_official_major_update_with_community_counterevidence; 最近核验：2026-07-23代码：Fixed commit 4d69d45cdc145d1ab5ec09f9872bfec542c05a06 exposes 4,119 tree entries and 534,583,596 blob bytes across asset generation, Kling/Blender HOI generation, 4D reconstruction, retargeting, tracking, export and visualization.
权重：Hub revision 943946a972d5de2eb0d2ff214b236d0e43575fd7 exposes pnp_ground, pnp_table and terrain release checkpoints with verified LFS identities and HTTP range access.
数据：The six currently listed categories sum to 22,312 motions and 5,576,500 frames. The card still reports 224 GB; advanced manipulation data and a turnkey quick-start remain TODO.
许可证：NVIDIA_noncommercial_code_plus_layered_dataset_terms；Root code is limited to noncommercial research/evaluation. GRAIL-original data is Apache-2.0; checkpoints, RoboCasa, ComAsset and per-mesh assets retain their own terms.
依赖：Docker and NVIDIA CUDA、Blender、Kling hosted video API、OpenAI API for a text/VLM stage、Hugging Face access、GEM-SMPL、FoundationPose、GMR、Isaac Lab、SONIC
硬件/传感器：Full paper training uses 64 NVIDIA L40 GPUs with 1,024 environments per GPU for about 30 hours per policy family、Real G1 deployment uses an external RTX 5090 and Luxonis OAK-D W at 10 Hz、At least 224 GB storage for the current dataset plus models and generated assets
实机证据：Seen-object pick-up is 84% and unseen-object pick-up is 80%, with ten trials per object.；Stair climbing is reported at 90%, but the paper does not expand its task-level denominator beyond the rate.
独立复现：No complete independent benchmark or G1 reproduction was found. GitHub issues document environment-specific native-extension failure on 8xH20/Python 3.10, a bottle-pickup behavior mismatch and currently unresolved missing GMR/table metadata paths.
限制/反面证据：The root code license is noncommercial and the dataset is legally layered by subtree.；The end-to-end route depends on hosted APIs and very large GPU/storage resources.；Open issue logs show that provided setup paths have required environment-specific rebuilds and still expose missing-path or tracking mismatches.；Physics-validated simulation trajectories are not physical safety evidence; the released corpus is single-embodiment G1 and lacks tactile/force labels.
当前建议：高潜待持续验证；Use the released data and task-general checkpoints for a bounded replay/tracking audit before attempting the full hosted-API pipeline; preserve the community setup failures as active counterevidence.
重点标签：人形合成数据工厂可能成为新主流、artifact:major-update、weights:task-general-released、dataset:22312-motions、license:layered、社区反证
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5 · 来源 6 · 来源 7 · 来源 8