official_author_supplementary_code · 已核验官方资源
审核码:verified_official; 官方关系:verified_official_author_repository_supplementary_implementation; 最近核验:2026-07-20代码:已核验开放。不可变提交 28b4516 包含 RobotLab 环境、定制 RSL-RL、MoE-CTS 训练、RoboGauge 接口、MuJoCo Sim2Sim 和 Go2 资产配置。
权重:仓库文档给出策略路径和最佳检查点评分,但没有核验到可直接下载、具独立许可证的论文策略检查点;需自行训练或使用官方主项目的其他发布仓库。
数据:不依赖离线训练数据集;地形和轨迹由 IsaacLab/RobotLab 环境生成。
许可证:Apache-2.0;Apache-2.0 覆盖该仓库源代码;Isaac Sim/Isaac Lab、RobotLab、RSL-RL、MuJoCo、Unitree SDK 与机器人资产分别适用上游条款。
依赖:Python 3.11、Isaac Lab 2.3.2.post1 / Isaac Sim and CUDA 12.8-compatible PyTorch 2.7.0、custom RobotLab 2.3.0 and RSL-RL 3.3.0 source trees、MuJoCo 3.4.0 or 3.6.0 and optional RoboGauge evaluation server
硬件/传感器:README benchmark example uses AMD EPYC 7763 and RTX 4090; evaluation takes about five minutes with 63 processes、Unitree Go2 for matching real deployment、repository documents Sim2Sim but does not provide a complete Unitree real-robot deployment command in this codebase
实机证据:Official paper/project reports Go2 deployment over snow, sand, stairs, slopes and 30 cm obstacles, plus high-speed tests up to 4 m/s.;This supplementary repository shows real-robot GIFs and publishes training/Sim2Sim code; physical deployment relies on additional project components and was not independently run.
独立复现:README reports RoboGauge score 0.6984 versus 0.6713 for the original go2_rl_gym port, but this is author-reported, not an independent reproduction; no qualifying third-party fixed-commit report was found.
限制/反面证据:这是同论文作者的补充 IsaacLab/RobotLab 实现,不应替代官方项目列出的 go2_rl_gym、RoboGauge 和 unitree_cpp_deploy 三个主仓库。;未核验到此仓库内可下载的论文策略权重。;真实 Go2 部署仍需要额外代码、SDK 和安全边界。;没有满足门槛的独立复现。
当前建议:高性能 Sim2Sim/评测候选;适合作为 IsaacLab 到 MuJoCo 的策略筛选和 MoE 足式控制高性能路线;先复现 RoboGauge 评测,再与主仓库和真机部署组件对齐,不能把作者内部比较当成独立横评。
重点标签:recency:2026、关联:主流足式控制技术2.0、关联:主流技术迭代、关联:近期热门、sim-to-real:predictive-evaluation、locomotion:mixture-of-experts、artifact:official-author-code、hardware:Unitree-Go2
核验来源:来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5