supermimic_vibed

supermimic_vibed 是一次「1 小时复现机器人学论文」的硬核实验：ETH Zurich 的 Chong Zhang 用 AI 编程助手在 1 小时内写出了 SuperMimic 论文的完整复现代码。这篇 CoRL 2026 论文（arXiv:2610.28480，SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model）提出用一个可微世界模型扩展人形机器人全身控制，但官方至今没有放出代码，作者于是按论文从零实现。

复现目标是论文的 LAFAN1 对照实验（图 4a）：把 40 段 LAFAN1 动作（舞蹈、跳跃、摔倒起身、冲刺等，共 2.45 小时）重定向到 29 自由度的 Unitree G1 人形机器人上，用单 GPU 训练一个「看一个动作、跟一个动作」的运动跟踪策略。整套代码跑在 Isaac Lab 3 的 kit-less 镜像上（MuJoCo-Warp 物理，无 PhysX、无 Kit），包含可微运动学、MLP(2048³) 世界模型、BPTT 策略优化——论文附录 A 的关键实现一个不少。

结果相当能打：在一块 RTX PRO 6000 上训练 1.47 小时后，评估归一化回报达到 0.829、40 个动作片段全部零失败，超过论文的对照目标（论文：H100 上 1.2 小时达到 0.82）。仓库里的 MILESTONES.md 记录了完整时间线：10 月 21 日 16:55 开工读论文，当晚 19:23 宣布复现成功。

作者在 README 的免责声明里写得直白，原话：「this repository was speed-vibecoded in 1 hour with an AI coding assistant.」还补了一句：「the code is too much "slop" for that: fast and unreviewed. Use it at your own risk.」——代码快但未经审查，风险自负。仓库建于 2026-10-21，不到一天就拿到 189★，MIT 开源。注意这不是开箱即玩的应用：复现需要本地 Linux + NVIDIA GPU + Docker 环境。

supermimic_vibed is a "reproduce-a-robotics-paper-in-an-hour" experiment: Chong Zhang (ETH Zurich) speed-vibecoded a full reproduction of the SuperMimic paper in 1 hour with an AI coding assistant. The CoRL 2026 paper (arXiv:2610.28480, "SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model") proposes scaling humanoid whole-body control with one differentiable world model — but no official code was released, so everything here is implemented from the paper.

The reproduction target is the paper's controlled LAFAN1 comparison (Fig. 4a): 40 LAFAN1 clips (dance, jumps, fall-and-get-up, sprint — 2.45 h total) retargeted to the 29-DoF Unitree G1 humanoid, training one motion-tracking policy on a single GPU. It runs on Isaac Lab 3's kit-less image (MuJoCo-Warp physics, no PhysX, no Kit) with differentiable kinematics, an MLP(2048,2048,2048) world model, and BPTT policy optimization — all the key ingredients from the paper's Appendix A.

The result holds up: after 1.47 GPU-hours on a single RTX PRO 6000, evaluation normalized return hit 0.829 with 0/40 failed clips, beating the paper's reference target (0.82 at 1.2 h on an H100). MILESTONES.md logs the full timeline: started reading the paper at 16:55 on Oct 21, declared reproduced at 19:23 the same evening.

The README disclaimer is refreshingly blunt, verbatim: "this repository was speed-vibecoded in 1 hour with an AI coding assistant." — followed by: "the code is too much \"slop\" for that: fast and unreviewed. Use it at your own risk." Created 2026-10-21, it collected 189 stars in under a day, MIT licensed. Note this is not a plug-and-play app: reproducing it needs a local Linux + NVIDIA GPU + Docker setup.