工程资源状态 代码 已核验可访问（使用限制见许可证） 权重 已核验可访问（使用限制见许可证） 数据 已核验可访问（使用限制见许可证） 许可证 许可证条款已核验（不等于可商用） 当前实践优先级 高潜待持续验证 · 76/100 · 0–6 个月
路线标签： 近期热门、可能成为新主流、地形感知人形全身控制技术2.0
社区证据类型： 尚无独立社区验证
把人类运动先验重定向为机器人中心地形约束并接入感知运动策略，固定策略仿真完成率由27.3%升至55.1%；代码、权重和99对G1动作开放，但真机仅有定性作者视频。
近期窗口 · 35/35发布距评分基准日 1 个月，落入 0–6 个月 窗口。 社区可复查验证 · 6/25官方工件可查，但截至核验日没有带实机指标的独立复现；按无显式社区证据封顶。 跨领域技术创新 · 20/20动作生成、形态重定向、地形几何优化与全身感知控制贯通。 量化实机证据 · 6/10G1实机视频存在，但量化试验表仅来自仿真。 开放与可复现 · 4/5代码、权重和动作可下载，混合BSD/CC-BY-NC条款限制复用。 工程实用增量 · 5/550Hz、29-DoF G1动作链和固定revision为工程验证提供明确起点。 社区可复查证据： 尚未收录；Stars、forks 与作者演示不计社区验证。
证据缺口： 缺少显式可复查社区复现/部署证据；无独立复现；真机无重复试验；BFM-Zero为非商用许可；48×A800训练成本；公开动作仅99对
已终审工程资源 官方代码 · 许可证 PMT-owned code BSD-3-Clause; bundled BFM-Zero vendor path is CC-BY-NC-4.0 and other third-party components retain their licenses · 最近核验 2026-07-22依赖：Python >=3.10、NVIDIA Isaac Sim and Isaac Lab、rsl_rl-compatible training stack、Unitree G1 assets provided separately、Git LFS or Hugging Face asset download、optional mjlab/MuJoCo-Warp backend for flat tasks；硬件/传感器：paper training uses 48 NVIDIA A800 GPUs for 10k iterations、Unitree G1 29-DoF、onboard depth sensing converted to height map、50 Hz policy control、CUDA workstation for Isaac Lab training and evaluation；实机证据：One shared policy is shown on a Unitree G1 across indoor, outdoor, stair, sparse-support, acrobatic and motion-capture mismatch settings.、The paper explicitly labels hardware evidence qualitative and reports no repeated-trial success or fall statistics.；开放范围/限制：The repository is a broad PMT research platform containing upstream/vendored methods; the paper result must be tied to its specific perceptive task and checkpoint. The repository as distributed is not unrestricted BSD because a vendored path is non-commercial. 官方数据集 · 许可证 Hugging Face dataset card BSD-3-Clause with upstream terms applying to bundled motion and model materials · 最近核验 2026-07-22依赖：Hugging Face CLI or Git LFS、PMT motion schema、NumPy NPZ reader、TCRS and terrain scene assets for regeneration；硬件/传感器：No robot is required to inspect the demo subset、Isaac Lab GPU environment for training、Unitree G1 and depth perception for real deployment；实机证据：The dataset is simulation/kinematic supervision used by a G1 policy with qualitative real transfer.；开放范围/限制：This is a 99-pair demonstration subset, not the full motion library or a complete reproduction snapshot for all 48-GPU paper training runs. 官方模型/权重 · 许可证 Hugging Face dataset card BSD-3-Clause; upstream/vendored checkpoint and model terms still apply · 最近核验 2026-07-22依赖：PMT fixed code revision、Isaac Lab task matching the selected checkpoint、Unitree G1 asset path、motion and terrain assets、checkpoint-specific provenance and normalization；硬件/传感器：CUDA GPU and Isaac Lab for most task playback、Unitree G1 for real transfer、flat-task subset can use the documented mjlab backend after compatibility work；实机证据：The perceptive policy lineage is deployed on G1 in official qualitative demonstrations; the downloadable bundle also contains checkpoints for other PMT-integrated methods.；开放范围/限制：The bundle is not a single homogeneous paper checkpoint license. Users must select the perceptive checkpoint and inspect its provenance; qualitative hardware videos do not establish a repeated-trial success rate. 每日候选人工终审 Perceptive BFM把人类平地动作先离线合成为地形一致参考，再以teacher-distill-PPO把原始动作命令与机器人侧高度图融合，使同一G1策略在楼梯、稀疏支撑、草地、室内外不规则地形保持动作风格。固定策略的跨地形仿真共450次rollout，完成率248/450（55.1%）对raw-reference 123/450（27.3%），下腿碰撞45/450对111/450。官方PMT仓库、预训练策略和99对raw/optimized示例真实开放；但48张A800训练、实机仅定性展示且明确没有重复试验，不能把视频覆盖当成功率。
实机证据： qualitative_single_policy_g1_no_repeated_trials；同一策略覆盖室内、室外、mocap mismatch、后空翻、舞蹈、横向/后退行走；论文明确因高能量动作安全风险只提供定性实机证据，没有重复试验统计。
硬件/本体： Unitree G1 29-DoF、onboard depth-to-height-map perception、50 Hz policy control
红旗与边界： 实机无重复trials或跌倒率；训练使用48张A800、10k iterations；TCRS假设静态刚性可观测高度场；上肢不做碰撞感知，论文展示撞障碍失败；仓库含CC-BY-NC vendored路径，不能整体视作宽松BSD
查看资源 · code · 状态 source_available_mixed_terms · 许可证 PMT-owned code BSD-3-Clause; vendored BFM-Zero path CC-BY-NC-4.0 and third-party components retain their terms固定提交包含342个Python/shell/YAML/TOML文件、Isaac Lab训练、TCRS、MuJoCo/mjlab适配、测试与约25个G1任务。 查看资源 · weights · 状态 verified_public_dataset_with_checkpoints · 许可证 dataset card BSD-3-Clause; individual vendored/checkpoint upstream terms still apply固定revision共221个文件约895,058,650 bytes，含15个预训练checkpoint/ONNX文件及SHA256SUMS。 查看资源 · data · 状态 verified_public_demo_subset · 许可证 dataset card BSD-3-Clause with upstream terms applying to bundled materials公开99对50fps、29-DoF G1 raw/terrain-optimized NPZ动作、地形网格与校验清单；不是完整论文训练动作库。 原始来源 