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

Nico Bohlinger, Jan PetersarXiv 2026 · arXiv

TL;DR

XMimic and XSoma are new frameworks that train humanoid robots for complex tasks—like dancing and navigating rough terrain—in under 60 seconds. By leveraging JAX-based compilation and a novel multi-GPU pipelining strategy, the authors achieve up to a 12x speedup over previous state-of-the-art methods, proving that model-based reinforcement learning can be both faster and more sample-efficient than model-free alternatives.

What Problem Does This Paper Address?

Training a humanoid robot to move like a human or walk over rubble typically takes hours or days of simulation time. While massively parallel GPU simulators (like Isaac Gym or MuJoCo) have brought this down to minutes, a significant bottleneck remains: sequential dependency.

Most modern "fast" methods rely on a long chain of small GPU operations. Even if you have a massive NVIDIA H100, the GPU spends more time waiting for the next instruction in the chain than actually performing heavy computation. Furthermore, model-based Reinforcement Learning (RL)—which learns a "world model" of physics to predict outcomes—has historically struggled to match the speed of model-free methods like PPO or SAC for locomotion tasks. The authors set out to break this sequential bottleneck and prove that world models are the key to sub-minute training.

Methods

The researchers introduced two systems built on the SuperMimic architecture:

  1. XMimic (for Motion Tracking): This system learns to follow pre-recorded human motion clips. It uses a learned world model to "imagine" short futures and optimizes the robot's policy using analytic gradients. Unlike previous versions, the entire training loop is compiled into XLA programs using JAX and the MuJoCo Warp (MJX) simulator.
  2. XSoma (for Locomotion): Designed for walking on rough terrain, XSoma takes a hybrid approach. It only uses the world model to learn the robot's internal dynamics. It computes external factors—like the exact height of the terrain under the feet and forward kinematics—using exact mathematical lookups. This allows it to learn with a one-second "imagination" horizon without needing a "critic" (a secondary network that estimates long-term value), which simplifies and accelerates training.

The Secret Sauce: Pipeline Parallelism To solve the GPU utilization problem, the authors restructured the training iteration. Instead of running data collection, model updates, and policy updates one after another, they run them simultaneously across up to four GPUs. While one GPU collects new data, another updates the world model, and a third optimizes the policy using a slightly "stale" version of the model. This "stale" data approach surprisingly has a negligible impact on learning stability but a massive impact on throughput.

Datasets

  • LAFAN1: A dataset containing 40 clips (roughly 2.5 hours) of human motion, including walking, dancing, and transitions, retargeted to the Unitree G1 humanoid.
  • Holosoma: A benchmark for humanoid locomotion on rough terrain, featuring tiles with pits, varying friction, and random "pushes" to test robot stability.

Key Insights and Ideas

  • Latency over Throughput: The primary bottleneck in fast RL isn't the total amount of math; it's the time it takes for a signal to pass through the "chain" of the world model during backpropagation. Shortening this chain is more effective than just adding more compute.
  • The "Stale" Advantage: In high-speed training, the policy and world model change incrementally. Using a model that is one or two iterations "old" to calculate gradients allows for massive parallelization without destabilizing the robot's learning.
  • Exact Math > Learned Math: For locomotion (XSoma), the authors found that learning the terrain height is a waste of time. By providing the policy with exact geometric clearances of the feet and body parts, the world model can focus entirely on the robot's complex joint dynamics.
  • Ditch the Critic: With a sufficiently long imagination horizon (1 second) and exact reward functions, the "critic" network—a staple of modern RL—becomes unnecessary for locomotion, removing a significant source of training noise and compute overhead.

Key Results

  • XMimic: Learned to track all 40 LAFAN1 clips on a Unitree G1 in just 0.86 minutes (51.6 seconds) using four H100 GPUs. This is 3.7x faster than the previous best, BlinkMimic.
  • XSoma: Reached the locomotion performance target in 0.77 minutes (46.2 seconds). This is 12.3x faster than FastSAC, the previous gold standard for fast locomotion.
  • Sample Efficiency: XSoma required only 1/8th of the environment steps compared to FastSAC to reach the same performance, highlighting the efficiency of analytic gradients over trial-and-error sampling.
  • Accuracy: After 15 minutes of training, XSoma achieved a velocity tracking error 33% lower than FastSAC.

Limitations and Future Work

  • Sim-to-Real: The paper focuses entirely on simulation. While the Unitree G1 is a real robot, the policies haven't been deployed on physical hardware yet, which involves dealing with sensor noise and latency not present in MJX.
  • Hardware Requirements: The sub-minute results require high-end hardware (H100 GPUs). While it runs on consumer cards like the RTX 4090, the scaling isn't as dramatic.
  • Terrain Complexity: While "rough terrain" was tested, the authors aim to extend XSoma to even more challenging environments, such as steep stairs or highly non-convex obstacles, in future work.

Summary updated

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.
BibTeX
@misc{bohlinger2026xmimic,
  author = {Nico Bohlinger and Jan Peters},
  title = {{XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute}},
  year = {2026},
  eprint = {2611.20452},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20452},
}
Figure