code_weights_and_data · 已核验官方资源
审核码：verified_official; 官方关系：verified_mit_real2sim_training_deployment_checkpoint_and_large_asset_release; 最近核验：2026-07-28代码：1,459 tracked files with real-to-sim, four-stage training and C++ deployment
权重：official Google Drive checkpoint folder plus scripted download in simulation path
数据：public nongated 33,895,302,407-byte Hugging Face assets.zip at fixed revision
许可证：MIT_with_upstream_terms；Root code and Hugging Face asset card are MIT; nested third-party code and separately hosted checkpoints retain their own or undisclosed terms.
依赖：Isaac Gym、PyTorch、ROS 1 Noetic、Unitree SDK2、Fast-LIO2、elevation mapping、MegaSaM、SAM2、WHAM、NKSR
硬件/传感器：two RTX 4090 GPUs for paper policy training、A100 for reported reconstruction timings、Unitree G1-23DoF with Jetson Orin NX and LiDAR
实机证据：Author qualitative single-policy G1 stairs, slopes and sit-stand deployment at 50 Hz; no repeated success denominator.
独立复现：mixed partial physical reproduction
限制/反面证据：Brittle monocular reconstruction；Rigid single-mesh scenes；Only 123 video clips；Coarse 11x11 LiDAR map；Preliminary deployment code；Independent G1 instability and embodiment sensitivity
当前建议：人类视频到情境全身控制主流技术2.0·分阶段验收；Checksum the 33.9 GB asset archive, reproduce one real-to-sim clip, one simulation checkpoint and a tethered flat-ground policy before stairs.
重点标签：video-to-humanoid、code:verified、weights:verified、data:verified、community:mixed-physical
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5