代码 · 已核验官方资源
审核码：verified_official; 官方关系：verified_mit_training_code_and_logs_without_pretrained_policy; 最近核验：2026-07-28代码：25 tracked files with HumanoidBench, MuJoCo Playground and IsaacLab paths
权重：not shipped
数据：official training logs shipped
许可证：MIT；MIT source with retained notices for adapted upstream code.
依赖：PyTorch、HumanoidBench、MuJoCo Playground fork、IsaacLab 4.5.0、Booster Gym fork
硬件/传感器：Paper uses one NVIDIA A100 and 16 CPU cores、default full-size humanoid setup can require 40-50 GB VRAM、Booster T1 for author physical deployment
实机证据：Author qualitative 12-DoF Booster T1 sim-to-real deployment; no physical trial denominator.
独立复现：mixed simulation reproduction
限制/反面证据：No official pretrained policy；Very high VRAM at default scale；Reward/action tuning is algorithm-specific；Final checkpoint sim-to-sim failure report；No independent physical reproduction
当前建议：高吞吐人形强化学习主流技术2.0·先复算仿真；Use training-log curves as a smoke-test target, store executed clamped actions consistently and validate an earlier checkpoint before the final checkpoint.
重点标签：off-policy-RL、code:verified、logs:verified、community:mixed
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4