SuperMimic

New
License: Apache License 2.0 · Updated 2026-11-12
My rating

1. Overview

SuperMimic is an open-source humanoid whole-body motion-tracking framework from the Hybrid Robotics Laboratory at UC Berkeley, released under the Apache 2.0 licence in 2026. It trains a controller that can follow large datasets of human motion by learning one differentiable world model of the robot's shared dynamics, rather than training a separate policy per motion or brute-forcing a single policy with cluster-scale compute. The work won the Best Paper Award at CoRL 2026 and was presented as an Oral. Its headline result is that a policy trained for three hours on a single GPU surpasses the return of a state-of-the-art motion-tracker baseline trained for seven days on one hundred twenty-eight GPUs, a reported reduction of more than one thousand times in GPU-hours. For a physics-AI programme the significance is that the simulator is no longer a passive data source but part of a learning loop closed through analytic gradients, and that measuring motion-tracking scaling laws becomes cheap enough to do on one machine.

2. How It Works

The core idea is that many different human motions share one body, and therefore one set of dynamics. SuperMimic first lets a policy run inside a standard rigid-body simulator and logs the resulting state transitions. It then fits a neural world model that, given the current state and an action, predicts the next state together with an immediate reward, the robot's observations and a fall-risk signal. The world model and the tracking policy are trained together in one loop. Because the world model is differentiable, backpropagation-through-time over short imagined rollouts gives the policy analytic gradients, and the policy is updated with the SHAC rule. Imagined rollouts inside the world model run more than ten times faster than stepping the simulator, and the first-order gradients provide a far stronger learning signal than the high-variance estimates used by model-free methods, so the policy reaches a target return in a fraction of the steps.

3. Scope and Assumptions

The framework targets whole-body motion tracking for humanoid robots, demonstrated on the Unitree G1, and it ships two task presets that differ in how the body is represented: generalized coordinates and maximal coordinates. It is built on mjlab, the MuJoCo-based research stack, and it depends on a standard rigid-body simulator to collect the transition data that the world model fits. The method assumes that all target motions share one underlying dynamics model, which holds for a single robot body but not across fundamentally different morphologies. It is a motion-tracking system, not a manipulation or locomotion planner, and it is demonstrated on humanoid bodies rather than quadrupeds, aerial vehicles or wheeled platforms. A GPU is required for training, and the reported results assume the G1 hardware and a large human-motion dataset; teams on other bodies should expect to re-collect data and re-fit the world model rather than reuse the published configuration.

4. Getting Started

The project uses the uv package manager. The instructions ask you to create a workspace directory, then clone both the supermimic repository and the mjlab_research repository as sibling folders, because SuperMimic imports training and playback code from mjlab. Two task identifiers are provided out of the box: SuperMimic-GenCoord-G1 for the generalized-coordinates representation and SuperMimic-MaxCoord-G1 for the maximal-coordinates representation. Training a world model and policy is a single command that takes a task identifier and a run name, and playing back a trained model takes the task identifier together with a Weights and Biases path. The fastest validation path is to train the generalized-coordinates task on the G1, confirm that the policy tracks a held-out motion in simulation, and only then consider the maximal-coordinates variant or a real-hardware deployment on the Unitree G1, so that failures can be attributed before any physical robot is involved.

5. Comparison with Alternatives

Against model-free reinforcement learning pipelines, the contrast is how the cost of learning is paid. A common route merges many per-motion specialist policies through lossy distillation, while another brute-forces one policy with cluster-scale compute; both make the GPU bill grow with the motion library. SuperMimic instead fits one world model of the shared dynamics and reuses it across all motions, so the data scale can grow while the per-motion training cost stays small. Against a physics engine used directly, SuperMimic consumes the simulator the way MuJoCo does, but it adds the learning layer on top and never replaces the engine. Against an upper-layer motion framework such as ProtoMotions, the relationship is complementary: SuperMimic focuses on efficient tracking-policy training through a world model, while broader frameworks organise tasks, assets and deployment. The win is efficiency, not feature breadth, and it is distinct from a GPU kernel substrate such as NVIDIA Warp.

6. Benchmarks and Evidence

The reported benchmark covers one hundred twenty-nine thousand five hundred fourteen motions. On a single NVIDIA H100 GPU and three hours of training, SuperMimic exceeds the return of the SONIC baseline, which was trained on one hundred twenty-eight NVIDIA L40 GPUs for roughly seven days, a reported reduction of more than one thousand times in GPU-hours. It fails on only twenty-five of the benchmark motions, versus one hundred seventy-nine failures for the baseline, and after nine point eight hours on two GPUs it reaches zero failures. The training data used is two hundred sixty-four point five six hours, about fifty-seven percent less than the baseline's six hundred eleven hours. The world model's eight-step prediction error stays between zero point zero zero seven and zero point zero one one as training data grows from one to two hundred sixty-four hours, and the policy transfers to a physical Unitree G1 across diverse and acrobatic motions. These figures are the project's own validation against a specific baseline.

7. Role in Physics-AI Workflows

SuperMimic sits at the point where a physics engine becomes a learning environment and a learned policy becomes a deployable controller. Its contribution to a physics-AI programme is to make the simulator interchangeable with a learned surrogate for the purpose of gradient-based training: the world model stands in for the rigid-body simulator during imagined rollouts, so policy improvement no longer requires the slow, high-variance sampling that model-free methods need. The shared-dynamics assumption turns the simulator from a per-motion cost centre into a one-time model-fitting step that amortizes across the whole motion library. For sim-to-real, the fact that a policy trained entirely inside the world model's imagination transfers to a physical G1 shows that the learned dynamics captured the relevant contact and balance behaviour. This is the kind of result that makes measuring motion-tracking scaling laws practical on a single machine rather than a compute proposal, and it sits alongside the physics-AI learning path and the physics-informed modelling branch.

8. Limitations and Failure Modes

The first limitation is the shared-dynamics assumption itself. It holds for one robot body, but a morphology change that alters the underlying dynamics can break the amortization and force re-collection and re-fitting. The second is dependence on simulator fidelity: the world model learns from logged transitions of a rigid-body simulator, so any contact or actuator artefact in that simulator is inherited, and the analytic gradients never see the true robot. The third is that the method is demonstrated on the Unitree G1 and a human-motion dataset; other bodies and datasets are not validated in the release. The fourth is the compute framing: the reported one-thousand-times reduction compares against a specific seven-day, one-hundred-twenty-eight-GPU baseline, and the practical saving depends on how closely a team's workload matches that baseline. The fifth is that tracking is not the same as task execution, and the framework does not by itself guarantee safe deployment on hardware.

9. Version and Source Notes

SuperMimic was released as open source in 2026 under the Apache 2.0 licence, with the code repository hosted on GitHub by the Hybrid Robotics Laboratory and a project page at supermimic.github.io. The accompanying paper, "SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model", was submitted to arXiv as 2610.28480 on 20 October 2026 and revised on 9 November 2026. It was accepted as an Oral at CoRL 2026 and received the conference Best Paper Award on 11 November 2026 in Austin, Texas. All seven authors are affiliated with the University of California, Berkeley; the senior authors are Pieter Abbeel and Koushil Sreenath. The code builds on mjlab, the MuJoCo-based research stack, and the repository acknowledges inspiration from ControlVAE, the robotic_world_model project and DreamerV3. The Apache 2.0 licence covers the source code; the paper itself follows the arXiv licence, and pretrained artifacts follow their own terms.

10. References

  1. SuperMimic repository README, GitHub, Hybrid Robotics, framework overview, task identifiers, training and playback commands.
  2. SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model, arXiv 2610.28480, submitted 20 October 2026, revised 9 November 2026, CoRL 2026 Oral and Best Paper.
  3. SuperMimic project page, supermimic.github.io, demonstration videos and visualisations.
  4. mjlab (MuJoCo Lab) research stack repository, the simulation foundation that SuperMimic's training and playback code depends on.
  5. MuJoCo entry, the standard rigid-body simulator and contact-dynamics solver that SuperMimic consumes.
  6. ProtoMotions entry, an upper-layer GPU-accelerated motion simulation and learning framework for humanoids and digital humans, complementary to SuperMimic.
  7. SONIC baseline coverage, the seven-day, one-hundred-twenty-eight-GPU motion-tracking baseline that the SuperMimic paper compares against.
  8. CoRL 2026, the robotics learning conference where SuperMimic was accepted as an Oral and won the Best Paper Award.
  9. DreamerV3 repository, differentiable world-model reinforcement learning work cited as inspiration in the SuperMimic repository acknowledgements.
  10. MuJoCo Playground, an open-source robot learning framework developed by the first author and winner of the RSS 2025 Best Demo Paper Award.
Tags: