XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute
Authors: Nico Bohlinger, Jan Peters
Organizations: Department of Computer Science, Technical University of Darmstadt, Germany. · Robotics Institute Germany (RIG); German Research Center for AI (DFKI); hessian.AI.
We present XMimic and XSoma, which learn humanoid motion tracking and locomotion on rough terrain in under a minute of training. Both follow SuperMimic, which learns a world model of the robot and optimizes the policy with analytic gradients through short rollouts imagined in this model. Existing implementations of this method do not make use of larger GPUs, as every training iteration is a long chain of small GPU operations that run one after another. XMimic and XSoma therefore shorten this chain and run the steps of every iteration at the same time on up to four GPUs. We compile the complete training loop with JAX, including the MuJoCo Warp simulator. For locomotion, XSoma leaves only the dynamics of the robot to the world model and computes everything else exactly, including the terrain height under the feet. It learns with a horizon of one second and without a critic. On four H100 GPUs, XMimic learns to track all 40 LAFAN1 clips with the Unitree G1 in 0.86 minutes, 3.7 times faster than BlinkMimic on one GPU. XSoma reaches the tracking target that Holosoma's tests set for FastSAC in 0.77 minutes, 12.3 times faster than FastSAC on one GPU, while needing only an eighth of its environment steps.
Figures & tables
Fig. 1: XMimic and XSoma train policies for the Unitree G1 in under a minute of training on four H100 GPUs. Top, an XMimic policy after 54 seconds of training, shown at five frames of a LAFAN1 dance clip. Bottom, an XSoma policy after 46 seconds of training, walking over Holosoma’s rough terrain while it follows velocity commands.
Fig. 2: Training iterations of XMimic (left) and XSoma (right) on one GPU and split across four GPUs, drawn to scale with times measured on H100 GPUs. On four GPUs, XMimic runs data collection, world-model update, and policy update at the same time on separate GPUs, while the fourth GPU runs the evaluations for the evaluation-guided resets. XSoma collects experience and updates the world model on one GPU, while three policy learners each compute a gradient and all of them apply the three gradients one after another.
Fig. 3: Motion tracking with XMimic on H100 GPUs. Time to the target (left), normalized return (middle), and number of failed clips (right) over the training time without compilation. The bars show the mean and the dots the individual runs, and the curves show the mean and standard deviation over the runs of each method. BlinkMimic has five runs, InstantTrack two, XMimic on one GPU eight, and XMimic on two to four GPUs four each. The dashed line marks the target return of 0.82.
Fig. 4: Locomotion on rough terrain with XSoma on H100 GPUs. Time to the target (left), Holosoma’s linear-velocity tracking reward, which is at most 2 (middle), and velocity error (right) over the first 15 minutes of training without compilation. The bars show the mean and the dots the individual runs, and the curves show the mean and standard deviation over the runs of each method. FastSAC and XSoma on four GPUs have eight runs each, XSoma on one GPU five, and the other methods four. The dashed line marks the target of 0.95.
Whole-body motion tracking lets a humanoid robot follow reference motions, but training a tracking policy for a library of motions has usually taken hours of GPU time. SuperMimic showed that the policy can instead be trained by backpropagating through a world model learned from parallel simulation, and InstantTrack added Muon updates, a residual policy around the reference and an averaged deployed policy. We build on both and present BlinkMimic, a recipe that makes this training faster. On a simulated Unitree G1, BlinkMimic trains one policy for all 40 LAFAN1 clips in 3.4 minutes on a single RTX 4090, measured until three consecutive checkpoints reach a normalized return of at least 0.82 with no failed clip (mean of three seeds). We split training time into the cost of an iteration, the iterations needed to fit the motions, and the iterations needed to remove the last failures, and assign each component, inherited or new, to one of these terms. The new components are a step size that decreases once training has progressed, resets directed by failures of the deployed policy under deployment conditions, and GPU graph capture of the whole iteration. A leave-one-out ablation with three seeds per component shows which term each component affects, and further measurements of the gradient, the update rule and the failure cases explain these effects. Code and per-seed results are available at https://github.com/fuyichen0725/blinkmimic.
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than 1000× and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.
We present SONIC-K1, to our knowledge the first open-weights whole-body motion tracker for the Booster K1 humanoid, obtained by adapting the SONIC framework to this smaller robot. The policy is trained from scratch on a single shared GPU with a small fraction of the compute reported for SONIC. In MuJoCo simulation, it tracks nearly all held-out motion clips and performs well on external benchmarks under SONIC's success criterion. The first release ran on the real K1 but shook its legs on steps and turns, and simulation showed a sensitivity to actuation delay. We therefore post-trained it with the robot's deployment torque limits, randomized actuation delay and gain randomization. The post-trained v1.0 stays robust under real-robot torque limits and large actuation delays, where the first release often failed, and in our runs on the real K1 it walks and turns without the earlier leg shaking. These results show broad motion-tracking capability on a modest training budget. Code, policy weights and evaluation tools are publicly available.
Kyrylo Kolesnichenko, Irvin Steve Cardenas, Jong-Hoon Kim
Institute of Social Sciences and Applied Informatics, Vilnius University, Lithuania · Advanced Telerobotics Research Lab - Computer Science Department, Kent State University, Kent, OH 44242, USA