[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2m1mpey9i4jjf":3},{"code":4,"message":5,"data":6,"timestamp":26},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":16,"summary_md":17,"topic":18,"github_topics":21,"license":22,"homepage":23,"updated_at":24,"history":25},"zita-ch\u002Fsupermimic_vibed",16,"vibe coded supermimic within 1h, LAFAN passed within 1.5 gpuh training","Python",475,42,144,3,[],"An unofficial reproduction of a humanoid whole-body control paper, for RL researchers who want LAFAN1 tracking without multi-GPU PPO.","## The problem\n\nHumanoid whole-body control on the LAFAN1 benchmark is normally a multi-GPU, multiple-day PPO job; the SuperMimic paper claims a differentiable world model changes that, but no official code exists. This repository reimplements the paper's controlled LAFAN1 experiment from scratch, ostensibly so one GPU and a coffee break are enough to see the result.\n\n## What's different\n\nThe usual route is PPO directly in a physics simulator: many parallel environments, long wall-clock, and cost that scales poorly when you add clips. Here the world model is learned and the policy is trained largely in imagination, with the simulator used to fill a replay buffer. Two things actually stand out: a single tracker covers all 40 clips, and the reported run reaches evaluation return 0.829 with zero failed clips in about 1.5 hours on one RTX PRO 6000 Blackwell — the README itself contrasts this with the paper's PPO baseline failing to reach 0.82 within 18 hours on up to 8 GPUs. Second, the stack runs kit-less: Isaac Lab 3 on Newton with the MuJoCo-Warp solver, no Isaac Sim\u002FKit and no PhysX, which is lighter than the standard Isaac Lab workflow.\n\n## Who it's for\n\nRL and robotics researchers who already know Isaac Lab, MuJoCo and the humanoid tracking literature, and who want to inspect or extend a world-model-based training loop rather than deploy a robot policy. It is not for people looking for production-grade locomotion code, sim-to-real results, or a clean library to build on — the author explicitly calls it unreviewed \"slop\", and self-collisions are disabled for speed.\n\n## Cost of adoption\n\nLinux and an NVIDIA GPU are mandatory; the setup expects Docker with the NVIDIA Container Toolkit, about 20 GB for images, and a specific Isaac Lab commit (46d6cc3). The data script pulls LAFAN1 G1 retargets under CC BY-NC-ND 4.0 from Hugging Face, so the motion data carries non-commercial, no-derivatives terms regardless of the repo's MIT license. Virtual-memory ulimits break Warp's CUDA-graph capture, and the asset conversion is offline because kit-less importers don't work in-process — expect to debug environment plumbing before you see a training curve. Commands are in the README rather than reproduced here.\n\n## Verdict\n\nThis is three days old with roughly 144 stars per day since creation, so it is a spike, not an established project, and the README admits it was written in an hour with an AI assistant. The reproduction target is concrete and the reported numbers line up with the paper's, which is more than most quick reproductions offer. But the code is unreviewed, the license on the data is restrictive, and no independent replication exists yet. Treat it as a readable reference implementation to study or fork, not as a dependency.",{"slug":19,"name":20},"machine-learning","Machine Learning",[],"MIT",null,"2026-10-25T02:05:11Z",[],1792909648]