UMR 0.1.0

This project will be continuously maintained and will subsequently evolve into a version implementation with a different focus from the original author's.

• Core Stages: Stage I learns reusable dense correspondences on a canonical T-pose; Stage II frame-by-frame solves it via Clarabel as a damped Gauss-Newton Quadratic Programming (QP) problem.
• Hardware Support: Comes with the Unitree G1 (29 DOFs) model out of the box.
• Program Entry Points: Two entry points share the same parameter set: retarget.py runs all three stages and launches the real-time player, while report.py generates metrics reports.
• Cache Optimization: Stage 0 and Stage I perform fingerprint caching under the skeleton signature and are shared across takes within the same shoot—requiring correspondences to be trained only once per batch.
• Performance: On the included 71-second walking test clip: without any manual bone mapping, it achieves 92.3% anatomical consistency, a median point error of 15.6 mm, 0.00 mm foot penetration into the ground, zero joint-limit violations, zero QP failures, and runs at 43.8 FPS.

The original author has open-sourced it at https://github.com/hanyang9/UMR.
Please check out their work and compare it with this repository for learning purposes. Thank you for your support! ❥(^_-)

---

该项目将持续维护，后续会演化出和原作者侧重点不同的的版本实现

• 核心阶段：第一阶段（Stage I）在规范的 T-pose（正向站姿）上学习可复用的稠密对应关系；第二阶段（Stage II）通过 Clarabel 将其作为阻尼高斯-牛顿二次规划（QP）问题逐帧求解。
• 硬件支持：仓库自带宇树 G1（29 自由度）模型，开箱即用。
• 程序入口：两个入口点共享同一组参数：retarget.py 运行全部三个阶段并打开实时播放器，report.py 生成指标报告。
• 缓存优化：阶段 0 和阶段 I 在骨骼签名下进行指纹缓存，并在同一场拍摄的各个片段之间共享，因此一批素材只需训练一次对应关系。
• 性能表现：在附带的 71 秒步行测试片段上：在没有任何人工骨骼映射的情况下，解剖学一致性达到 92.3%，中位点误差为 15.6 毫米，足部穿地为 0.00 毫米，零关节极限违规，零 QP 失败，运行帧率为 43.8 FPS。

原作者已经开源https://github.com/hanyang9/UMR
希望大家关注他们，也可以和本仓库对比学习 ，感谢关注 ❥(^_-)