code_and_weights · 已核验官方资源
审核码：verified_official; 官方关系：verified_apache_training_evaluation_and_tracked_pt_onnx_without_physical_deployment_scripts; 最近核验：2026-07-29代码：995 tracked entries with teacher PPO, student DAgger, evaluation and ONNX export paths
权重：teacher/student PT and ONNX files are tracked under logs_rl
数据：no paper-scale real-door dataset
许可证：Apache-2.0；Root source and tracked files are published under Apache-2.0 subject to retained third-party asset terms.
依赖：Ubuntu 22.04、Python 3.11、NVIDIA driver >=535、Isaac Sim 5.1、Isaac Lab、PyTorch 2.7、CUDA 12.8
硬件/传感器：CUDA GPU for training、paper physical setup uses Unitree G1, two 7-DoF three-finger hands, Intel RealSense D435i and RTX 4090 desktop
实机证据：Author reports 83% aggregate door success and 15.40 s completion but does not disclose physical trial count.
独立复现：partial_simulation_success_without_task_completion
限制/反面证据：No public physical deployment scripts；System-identification details incomplete；Physical success denominator missing；Independent run opens the door but does not traverse；No independent physical reproduction
当前建议：RGB人形门操作颠覆性分支·先复算仿真闭环；Reproduce one held-out simulation door including threshold crossing, then implement a separately reviewed G1 safety/deployment layer.
重点标签：sim-to-real、pixel-to-action、code:verified、weights:verified、community:partial-sim
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4