SuperMimic: Humanoid Motion Tracking With One World Model, and What It Means for Our G1 Pipeline

SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model (arXiv:2610.28480, 20 October 2026). Authors: Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel and Koushil Sreenath; UC Berkeley. Shao and Long contributed equally; Abbeel and Sreenath advised equally. According to the arXiv record, the work was accepted at CoRL 2026 as an oral presentation.

The topic is whole-body motion tracking: the robot learns to replay a reference motion derived from human motion capture with its own joints, without falling and while keeping balance. The work builds on two earlier efforts. The tracking reward comes from BeyondMimic, whose first author is also Qiayuan Liao. The point of comparison is NVIDIA's large-scale tracking policy SONIC. The target robot is the Unitree G1, shown both in simulation and on hardware.

In the classic approach a policy learns from millions of trials in a physics simulator (usually with PPO). SuperMimic inserts a learned world model and turns training into a three-step loop:

• Collect: the policy runs in a standard contact simulator (MuJoCo Warp); experience goes into a replay buffer.
• Fit: one world model for all motions is fitted to that data and updated every iteration.
• Imagine: the policy is trained on short 32-step imagined rollouts inside the world model. Because the model is differentiable, the policy gradient is computed directly by backpropagation through time (BPTT, an analytic policy gradient).

One key detail: gradients never pass through the simulator. The simulator only collects data; derivatives come from the learned model, which is held fixed inside each policy update. A critic bootstraps beyond the horizon; according to the authors, 16 steps with a critic match 32 steps without one, and without a critic training fails at 16 steps or fewer. One model step equals one 50 Hz control step, that is, four 5 ms MuJoCo Warp substeps.

The authors' speed claim: on an H100 the world model is more than 10x faster than MuJoCo Warp at practical environment counts (11x at 2k environments, 26x at 16k, 62x at 64k). In an ablation on walk, run and sprint motions from LAFAN1, first-order gradients reportedly reach the best training return of a PPO variant on the same world model with about 64x fewer imagined steps. The world model's normalized 8-step prediction error stays within 0.007-0.011 from 1 to 264 hours of data.

Evaluation uses 129,514 motions (264.56 hours) from the filtered training subset of Bones Studio's BONES-SEED dataset. Each motion plays once from the first frame; foot friction is 1.0, with no domain randomization, no observation noise and deterministic physics. A motion counts as failed if it triggers SuperMimic's own training termination condition. Failed motions as reported in the paper:

The authors' compute comparison:

From this the authors claim more than 1000x fewer GPU-hours (about 7200x by the authors' accounting) and more than 50x less wall-clock time (about 56x). The 3 hours cover the full loop (collect, fit, policy update) and exclude checkpoint evaluations. On hardware, a single policy reportedly runs on a physical G1 zero-shot from simulation with no per-motion tuning; the project page shows video clips of a side roll, a 180 degree flip, a roll in a forest and a 28-take montage of dance, rolls and jumps, and Fig. 6 of the paper adds frame sequences of a jump, a side roll, a 90 degree flip and an in-place 360 degree flip.

The numbers are impressive, but it matters what they measure.

• Different hardware. The GPU-hour ratio counts GPUs of different generations. The authors say so themselves: an H100 has 0.74x the fp32 peak and 5.5x the tensor-core peak of an L40. The SONIC paper does not name the GPU model; the L40 detail comes from the SuperMimic paper. The current version of the SONIC paper (v4) gives training as 128 GPUs over 7 days and about 21,000 GPU-hours.
• Different data. SONIC was trained on a 611-hour motion set; SuperMimic on a filtered 264.56-hour portion of the public BONES-SEED set (2.3x less data, per the authors). The evaluation runs on that subset.
• Benchmark = training set. The evaluated motions are in both methods' training data. The authors state this plainly: what is measured is coverage, how well trained motions are reproduced, not generalization to motions never seen.
• SuperMimic's criterion. Failure is defined by SuperMimic's own termination condition, and SONIC is judged by it. The 0.90 score and 179 failures attributed to SONIC are the SuperMimic team's measurements.
• Easy-side conditions. No randomization, no observation noise, deterministic physics. On hardware there are only video clips, no per-motion success rate.

None of this invalidates the result: very high coverage on the same dataset with far less compute is a strong signal for the world-model approach. But "7200x more efficient than SONIC" should not be read as two systems doing the same job under the same conditions.

At Robotlar.org we run our own motion tracking pipeline for dance shows on the Unitree G1: human motion is extracted from video (GVHMR), retargeted to the G1's 29-DOF skeleton (GMR) and trained with the BeyondMimic tracking formulation in mjlab. Since SuperMimic uses a modified BeyondMimic tracking reward, the comparison is relevant. The run times below come from our own training logs; the SuperMimic figures are the results reported in the paper.

• Today every dance is a separate policy. Each dance needs its own training run, its own ONNX file and its own entry on the deploy side. The first 30,000-iteration run for one dance takes about 10.5 to 13 hours on a 12 GB laptop GPU (RTX 5070 Ti Laptop); the second pass after cleaning the noisy video-derived reference takes about 7 to 8 hours. Run time is set by the fixed iteration count, not by dance length: a 17-second and an 81-second dance took similar time.
• The likely gain is not training time but the fixed cost per dance. If the claims hold, one general policy could remove the separate training run, ONNX management and first cleaning pass for each dance. But since the benchmark measures the training set, the paper does not measure whether a new zeybek extracted from video could be tracked with no extra training. The likely route is adding the new dance to the data and training a little longer; the paper does not report that cost.
• Our real bottleneck is reference quality, not GPUs. Video-derived motion has artifacts such as foot sliding and ground penetration. A tracker rollout in physics cleans up ground penetration but probably does not fully remove foot sliding; a better tracker alone may not fix a sliding reference.
• Approval on the robot does not go away. We approve every dance on the real robot one by one. A ready-made dance policy from outside, not approved on the robot, tripped the safety protection and went into a protected collapse on the real robot. A single general policy still has to be approved motion by motion on hardware.
• Hardware and integration. Backpropagating 16-32 steps through a world model probably needs more activation memory than PPO at the same batch size; per the paper, 128 steps is the longest horizon that fits on an H100. Full-scale training is not expected to fit on a 12 GB local GPU. Our current dance profile expects the motion baked into the ONNX; a general tracker needs a separate policy and config definition plus sim-to-sim testing. With no code or weights linked, rebuilding the world model and BPTT pipeline from scratch is not a configuration change but a research project in its own right.
• Flips and rolls. This is where SuperMimic's gap is most striking. Our repertoire has no flips or rolls; motions where the robot deliberately goes to the ground mean separate hardware risk and separate safety approval.

• Code and weights. The project page, the arXiv record and the paper contain no code or weights link.
• The web demo is proprietary. The demo on the project page runs the policy in the browser, but its licence reads "all rights reserved": only personal, non-commercial evaluation is allowed; copying, downloading and storing, reverse engineering, extracting parameters and training other models on the demo's outputs are prohibited. The demo model cannot be used in a product.
• Paper licence. The arXiv record uses the standard distribution licence (nonexclusive-distrib 1.0), not Creative Commons.
• BONES-SEED licence. The training data is Bones Studio's BONES-SEED set (142,220 motions, about 288 hours, 522 performers). Access opens on accepting a licence. The licence grants rights to academic users and to startups with annual revenue under USD 1 million; other commercial use needs a separate licence. Model training is allowed, raw data may not be redistributed, and trained models must credit Bones Studio.
• LAFAN1. LAFAN1, used in the paper's ablations, the PPO-WM comparison and the training of the velocity interface, is licensed CC BY-NC-ND 4.0. Commercial use of a policy trained on it carries licence risk.
• Limited real-world evidence. Hardware results are videos and frame sequences, not numeric rates; evaluation conditions have no randomization and no noise.
• Generalization unmeasured. How well a motion outside the training set would be tracked is an open question.

If you have a working per-dance policy pipeline, keep it; SuperMimic is not software you can try today but a report of results. For a team that wants to try a general tracking policy on the G1 today, the released option is SONIC: NVIDIA published G1 checkpoints on Hugging Face under the NVIDIA Open Model License and the code under Apache-2.0. For per-dance training, the BeyondMimic formulation lives on in mjlab (Apache-2.0).

If SuperMimic's code or weights are released, the order we suggest: first read each licence separately (training data, checkpoint, code); then run a sim-to-sim comparison in simulation with your own reference motions against your existing per-dance policies; only after that move to the robot, with a safety gantry. Plan safety separately for ground-contact motions such as flips and rolls.

On the hardware side, the G1 EDU is the target platform for this line of research. For simulation and training infrastructure see our RL Gym and G1 research platform pages. For university procurement through BAP and TÜBİTAK, see our institutional buying guide.

SuperMimic is an independent academic work by researchers at UC Berkeley (Qiayuan Liao, Yiyang Shao, Junfeng Long, Kevin Zakka, Roberto Horowitz, Pieter Abbeel, Koushil Sreenath); Robotlar.org has no institutional affiliation with the work or its authors. The values in the tables are the results reported in the paper. The training times for the G1 pipeline are based on Robotlar.org's own training logs. Robotlar.org is the Turkey partner of Unitree Robotics; contact us about motion tracking and research configuration on the G1. Sources: the project page, arXiv:2610.28480, SONIC (arXiv:2511.07820) and BeyondMimic (arXiv:2508.08241).