← Paper Digest

LocoMuJoCo: A Comprehensive Imitation Learning Benchmark for Whole-Body Control

Firas Al-Hafez, Davide Tateo, Oleg Kaidanov, Aditya Bhatt, Artur Zelik, Zhengyi Luo, Jan Peters

November 2026

Simulated humanoids tracking reference motions, then many humanoids and quadrupeds in parallel environments.

TL;DR An open-source MuJoCo/JAX benchmark for whole-body imitation and RL with 16 robot and human embodiments, motion-capture datasets, automatic retargeting and a sim-to-real path. Its GPU backend reaches 4.9×10⁵ steps per second on one consumer GPU, and a policy trained in it runs a 43 s dance on a real AgiBot X2 Ultra.

What they do

Humanoid imitation and RL research is spread across single-robot codebases, so changing robots means rebuilding data pipelines, retargeting, simulators and metrics. LocoMuJoCo puts all of these behind one API: 9 humanoids (Unitree H1/H1v2/G1, Atlas, Talos, Apollo, Booster T1, ToddlerBot, Fourier GR1T2), 3 SMPL-derived skeletal and muscle models, and 4 quadrupeds, with mocap data and a companion repository, loco-mujoco-real, for deploying policies on hardware.

How it works

  • Backends. Environment code is written once and runs on CPU MuJoCo, MJX or MjWarp. Observations, rewards, control modes, domain randomization and terrain are stateless plug-ins, so the whole step is JIT-compilable.
  • Retargeting. An SMPL body shape is fitted to each robot once using 15 mimic sites; human clips (native dataset, LAFAN1, AMASS) are then tracked with soft MuJoCo constraints that respect collisions and joint limits. Robot-to-robot transfer goes back through SMPL.
  • Algorithms and metrics. JAX implementations of PPO, GAIL, AMP, DeepMimic-style tracking, SuperMimic, SAC, TD3 and others, plus Euclidean, DTW and Fréchet trajectory metrics.

Results

  • On a Unitree G1 with full self-collision, MjWarp reaches 4.9×10⁵ steps/s on one RTX 4090, 20-54× more than MJX, and runs 1.25-1.9× faster than mjlab at the environment level.
  • On Unitree H1 over 13 clips, PPO reaches a return of about 1530 vs. 1260 for SAC and 1080 for TD3. PPO + DeepMimic beats GAIL by about 2× in return and 5-10× in joint-position error.
  • One SuperMimic policy per robot on all 40 LAFAN1 clips (about 50 min on one GPU) completes 99.7% of frames on the X2 Ultra and 99.1% on the G1 in simulation.
  • A SuperMimic policy trained on all of BONES-SEED runs a 43 s chain of its dance clips on the real X2 Ultra at 50 Hz using onboard sensing only.

Limitations

The paper has no limitations section. Sim-to-real tooling covers two robots, and the hardware run is shown as a figure without metrics. Algorithm comparisons use one humanoid and 13 clips, and the LAFAN1 evaluation turns randomization off, so it measures tracking rather than robustness.

Figures

From Figure 1: The 12 humanoid and (musculo-)skeletal embodiments replaying one walking trajectory, with ghosted pose...
From Figure 1: The 12 humanoid and (musculo-)skeletal embodiments replaying one walking trajectory, with ghosted poses showing time.
From Figure 2: One motion-tracking policy trained in LocoMuJoCo running on the real AgiBot X2 Ultra over a 43 s motion.
From Figure 2: One motion-tracking policy trained in LocoMuJoCo running on the real AgiBot X2 Ultra over a 43 s motion.
From Figure 4: The three retargeting stages. (a, e) SMPL shape fitted to the Unitree G1 before and after optimization...
From Figure 4: The three retargeting stages. (a, e) SMPL shape fitted to the Unitree G1 before and after optimization, mean 15-site error falling from 112.1 to 21.8 mm. (b-d) A dance clip retargeted from SMPL onto a Fourier GR1T2. (f-h) A Unitree H1v2 walk transferred to a Booster T1 through a synthetic SMPL intermediate.
From Figure 6: PPO + DeepMimic on 11 embodiments, 13 clips and 5 seeds. Top: training return and episode length per e...
From Figure 6: PPO + DeepMimic on 11 embodiments, 13 clips and 5 seeds. Top: training return and episode length per embodiment. Bottom: final joint-position errors (Euclidean, DTW and Fréchet; log scale, lower is better).
Abstract

LocoMuJoCo is an open-source benchmark for imitation learning (IL) and reinforcement learning (RL) on whole-body control. The library couples MuJoCo (CPU) with MJX and the MjWarp backend (GPU) under a single API, exposes 9 humanoid, 3 (musculo-)skeletal, and 4 quadruped embodiments, and ships a curated native motion-capture dataset for every humanoid, to which it can also retarget AMASS and LAFAN1. A modular plug-in architecture makes observations, rewards, control modes, terrains, and domain randomization fully interchangeable, while compact JAX implementations of many RL and IL algorithms provide JIT-compiled end-to-end training, and the companion loco-mujoco-real repository provides sim-to-real deployment. We describe the design, the SMPL-to-robot and robot-to-robot retargeting pipeline, and a usage example. LocoMuJoCo is released under MIT at https://github.com/robfiras/loco-mujoco.

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.