official_cross_embodiment_training_code_partial_release · verified_official_code_partial_release
审核码：verified_official_code_partial_release; 官方关系：verified_official_repository_code; 最近核验：2026-07-20代码：已核验开放。提交 f2d91d5 包含 74 个 Python 文件、动作重定向、强化学习/模仿学习管线与部分机器人配置。
权重：README roadmap 仍把论文检查点列为计划项；Issue #5 继续请求预训练策略，未发现正式 Release。
数据：未发布论文完整训练轨迹或统一跨本体数据集；输入动作和机器人模型依赖各自来源。
许可证：Apache-2.0；根 Apache-2.0 覆盖仓库发布代码。Unitree、Fourier、Turin 机器人模型、动作数据、Isaac Gym/Lab 与第三方基础代码仍按各自条款使用。
依赖：Python/PyTorch training stack、Isaac-based simulation and robot-specific URDF/MJCF assets、motion retargeting plus behavior-cloning RL/IL stages、MuJoCo sim-to-sim and physical deployment paths are not yet public
硬件/传感器：README reports retargeting on RTX 3070 Ti 8 GB、RL/IL reported on RTX 4090D 24 GB or RTX 5090D 32 GB and recommends at least 24 GB、paper tests Unitree G1, H1-2, Fourier GR1 and Turin V3 in simulation only
实机证据：The paper limitation explicitly states that evaluation was conducted in simulation and lacks physical deployment.；The public roadmap confirms that deployment code is not yet released.
独立复现：Issue #4 documents an independent retargeting run with modified URDF, mappings, weights and video, but reports foot sinking versus GMR. The author attributes this to virtual-toe/SMPL foot-height mismatch and recommends re-fitting or a downward offset; no paper-comparable metric or fixed resolved commit is supplied.
限制/反面证据：论文四种本体只在仿真评测，尚无真机控制证据。；检查点、所有机器人训练脚本、MuJoCo sim-to-sim 和部署代码仍在 roadmap。；社区迁移暴露虚拟脚趾与 SMPL 脚高不匹配，跨本体并非零配置。；训练阶段建议至少 24 GB 显存，不属于极低成本训练路线。
当前建议：主流跨本体模仿技术 2.0 候选，部分开放；统一 retargeting 与 RL/IL 的跨本体接口有较高工程杠杆，且社区已经暴露具体迁移误差；当前适合做 retargeting 最小验证，等待完整训练与真机栈后再判断能否成为新主流。
重点标签：recency:2025-paper-2026-code、关联:主流全身控制技术2.0、关联:近期热门、关联:社区验证、关联:技术改进版、cross-embodiment:four-humanoids、artifact:code-open-checkpoints-deploy-pending、evidence:simulation-only
核验来源：来源 1 · 来源 2 · 来源 3 · 来源 4 · 来源 5