CuboiLeo/SuperMimic-Reproduced

What it solves

This project provides a from-scratch, independent reproduction of the "SuperMimic" research paper, which focuses on training humanoid robots to perform complex, whole-body movements. It addresses the challenge of scaling humanoid control by using a differentiable world model, allowing the robot to learn diverse motions—such as jumping, spinning, and crawling—directly from large motion datasets.

How it works

The system uses a differentiable world model to simulate the robot's environment and predict future states. It trains a policy using Backpropagation Through Time (BPTT), which allows the robot to learn by "imagining" sequences of actions and optimizing them against a reference motion library. The implementation is built on MuJoCo Warp and mjlab, utilizing a custom motion library, differentiable kinematics, and data-parallel training to achieve high performance on the Unitree G1 robot.

Who it’s for

This project is for researchers and robotics engineers interested in humanoid whole-body control, differentiable simulation, and model-based reinforcement learning. It serves as a reference implementation for those looking to replicate the results of the SuperMimic paper or build upon its specific training protocols and world-model architecture.

Highlights

• High-Performance Reproduction: Achieves a 0.930 normalized return on the BONES-SEED dataset, closely matching the paper's performance.
• Full Pipeline Implementation: Includes everything from data processing and motion library building to the differentiable world model, BPTT trainer, and evaluation benchmarks.
• Optimized Training: Supports data-parallel training across multiple GPUs, reaching high simulation speeds (approx. 2×10⁵ env-steps/s) on an RTX 5090.
• Transparent Documentation: Provides detailed logs, design decisions, and a breakdown of how the code maps to specific sections of the original research paper.