zita-ch/supermimic_vibed

What it solves

Controlling a humanoid robot to imitate complex whole-body motions (like dance, jumps, fall-and-get-up, sprint) is hard. The SuperMimic paper showed that a differentiable world model can make training far more efficient, but its official code was never released. This project independently reimplements the method and shows it works: it matches the paper's accuracy in under 1.5 GPU-hours on a single NVIDIA card.

How it works

The system runs on Isaac Lab 3 with a "kit-less" build that only uses the Newton physics engine with the MuJoCo-Warp solver — no Isaac Sim, no PhysX. It uses a 29-degree-of-freedom G1 humanoid model and 40 LAFAN1 motion clips (dance, jumps, etc.) retargeted to that robot.

Training is based on SuperMimic's recipe: a motion encoder turns the reference motion into a context vector, then a policy and a learned world model operate together. The world model predicts future states of the robot, and the policy is trained by "imagination" — it plans many steps ahead inside that model and propagates gradients back through time. The setup uses 4096 parallel environments, return-based reset sampling, and a set of hyperparameters that match the paper.

The project also ships Dockerfiles, data download scripts, and tools to convert the robot model from URDF to MJCF and USD, so the whole pipeline is reproducible.

Who it's for

Anyone working on humanoid control, model-based RL, or reproducing state-of-the-art papers without official code. It's also a useful reference for running Isaac Lab without the full Kit environment.

Highlights

• Matches the paper's evaluation return: 0.829 vs 0.82, with 0/40 failed clips
• Requires only ~1.5 GPU-hours on a single RTX PRO 6000 (paper used 1.2h on an H100)
• Kit-less: no Isaac Sim, uses Newton + MuJoCo-Warp
• Includes Docker setup and data download scripts for full reproducibility
• Provides evaluation watcher, plotting, and rendering tools
• Code is MIT licensed; motion data is not redistributed (CC BY-NC-ND)