Awesome SuperMimic — A Curated Resource List

This repository is a community-maintained list of projects related to SuperMimic, a method for controlling humanoid robots. It collects implementations, extensions, demos, and other resources, all linked from a single page.

What is SuperMimic?

SuperMimic is a research method (from UC Berkeley, presented at CoRL 2026) that teaches a humanoid robot to imitate complex motions (like flips, dances, and rolls). It works by:

• Learning a world model — a simulation of how the robot moves — from short snippets of training data.
• Using that model to imagine short future rollouts and improve a motion-tracking policy.
• Training the world model and policy together in one loop, so the policy learns through the model.

The original paper, video, and an interactive browser demo are linked from the README.

What This List Contains

The README organizes 38 community projects into categories:

• Real-robot deployment and Sim2Real (2 projects) — run the policy on actual G1 robots.
• Independent reproductions (5 projects) — reimplement SuperMimic from scratch, with training and evaluation notes.
• Method and morphology variants (5 projects) — change the algorithm or adapt it to different robot bodies.
• Perception, terrain and quadrupeds (4 projects) — add vision, height scans, or quadruped tasks.
• Manipulation and dexterous hands (4 projects) — arm, hand, and object-interaction tasks.
• Muscle, interaction and diagnostics (3 projects) — muscle-actuated bodies, physical interaction, contact studies.
• Frameworks and training stacks (6 projects) — reusable training or evaluation pipelines.
• Robot porting and retargeting (6 projects) — ports to specific robots, simulators, or hardware.
• Inspired experiments (3 projects) — apply SuperMimic ideas to other control problems.

Each entry links to the project's repository, often with a quickstart guide, results, and a video or GIF.

Key Details and Caveats

• Official code status: As of the README's last check (2026-10-28), the official SuperMimic website does not link a training-code repository. All listed projects are community implementations or extensions.
• Results are upstream reports: The README explicitly states that numbers, videos, and demos come from the linked projects themselves; this list has not independently rerun the experiments.
• Some entries were recently refreshed (12 on 2026-10-28) against upstream documentation and source files; others retain earlier snapshots.
• Hardware and data differences: Many projects note that their training times and results are specific to their GPUs (e.g., RTX 5080, A6000) and datasets (LAFAN1, BONES-SEED), and may not match the original paper's H100-based numbers.

How to Use This List

The README is a table of contents with anchors. You can jump directly to a category (e.g., "Real-robot deployment") or browse the full list. Each project entry includes a short description, upstream-reported results, and a link to the source repository. The list also includes a citation section and a star-history chart.

In short: if you want to find or compare SuperMimic implementations, this is the go-to index. It's a curated, community-driven directory — not a codebase itself.