← Paper Digest

XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute

Nico Bohlinger, Jan Peters

November 2026

TL;DR XMimic and XSoma train Unitree G1 motion-tracking and rough-terrain walking policies through a learned world model in under a minute, by compiling the whole training loop in JAX and splitting each iteration across four H100 GPUs. Tracking all 40 LAFAN1 clips takes 0.86 min (3.7x faster than BlinkMimic) and walking reaches its target in 0.77 min (12.3x faster than FastSAC).

What they do

SuperMimic-style training optimizes a humanoid policy with analytic gradients through short rollouts in a learned world model, which needs far less compute than PPO. Existing implementations barely speed up on bigger GPUs, because each iteration is a long chain of small dependent GPU operations. XMimic (motion tracking) and XSoma (velocity tracking on rough terrain) shorten that chain and spread each iteration over up to four GPUs. All training and evaluation is in simulation on the Unitree G1; code for both is released.

How it works

  • Compiled loop. The whole iteration, including the MuJoCo Warp simulator, is compiled with JAX. Work that does not depend on imagined states runs once before the rollout, and the backward pass uses fewer, larger bfloat16 matrix multiplications, making the policy update almost 40% faster than a direct port of BlinkMimic.
  • XMimic pipeline. Data collection, world-model update and policy update run at the same time on three GPUs, using a policy and world model one or two iterations old; a fourth GPU runs evaluations.
  • XSoma. Policy updates take about 70% of an iteration, so several GPUs compute gradients of consecutive updates in parallel. The world model predicts only robot dynamics; rewards, falls and terrain height under the feet are computed exactly, with a 48-step (about 1 s) horizon and no critic.

Results

  • Tracking all 40 LAFAN1 clips: XMimic reaches BlinkMimic’s target in 0.86 min on four H100s, 3.7x faster than BlinkMimic on one, and is 1.3x faster on a single GPU. Stale updates cost only about 10% more iterations.
  • Locomotion: XSoma reaches the FastSAC target from Holosoma’s tests in 2.50 min on one GPU (3.8x faster than FastSAC) and 0.77 min on four (12.3x), using an eighth of FastSAC’s environment steps. PPO misses the target within an hour.
  • After 15 min, XSoma’s velocity error is a third lower than FastSAC’s.

Limitations

No real-robot deployment yet; the terrain has height differences of only a few centimeters. On four GPUs data collection becomes the bottleneck, and reported times exclude 20 s to 3 min of start-up and compilation.

Figures

From Figure 1: An XMimic policy for the Unitree G1 after 54 seconds of training on four H100 GPUs, shown at five fram...
From Figure 1: An XMimic policy for the Unitree G1 after 54 seconds of training on four H100 GPUs, shown at five frames of a LAFAN1 dance clip.
From Figure 2: Training iterations of XMimic (left) and XSoma (right) on one GPU and split across four, drawn to scal...
From Figure 2: Training iterations of XMimic (left) and XSoma (right) on one GPU and split across four, drawn to scale with H100 times. XMimic runs data collection, world-model update and policy update on separate GPUs, with evaluations on the fourth; XSoma collects data and updates the world model on one GPU while three policy learners each compute a gradient and all apply the three gradients in turn.
From Figure 3: Motion tracking with XMimic on one to four H100 GPUs against BlinkMimic and InstantTrack: time to the ...
From Figure 3: Motion tracking with XMimic on one to four H100 GPUs against BlinkMimic and InstantTrack: time to the target (left), normalized return (middle) and number of failed clips (right) over training time. The dashed line marks the target return of 0.82.
From Figure 4: Rough-terrain locomotion with XSoma on one to four H100 GPUs against FastSAC and PPO: time to the targ...
From Figure 4: Rough-terrain locomotion with XSoma on one to four H100 GPUs against FastSAC and PPO: time to the target (left), linear-velocity tracking reward (middle) and velocity error (right) over the first 15 minutes of training. The dashed line marks the target of 0.95.
Abstract

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.