Skip to main content
Back to timeline
arXivSource publication:

LocoMuJoCo moves whole-body control benchmarking onto the GPU: one API, and MjWarp reaches 487,553 environment steps per second at 8,192 parallel environments

Synopsis

LocoMuJoCo is released as an open-source library that unifies imitation-learning and reinforcement-learning benchmarking for whole-body control under one API: the same environment code runs on a serial NumPy/MuJoCo backend or a parallel JAX/MJX or MjWarp backend, covering 9 humanoids, 3 (musculo-)skeletal models and 4 quadrupeds, each humanoid with a native motion-capture dataset. Observations, rewards, control modes, terrains and domain randomization are interchangeable plug-ins, and the library adds automatic SMPL-to-robot retargeting plus the loco-mujoco-real deployment pipeline. On the Unitree G1, the MjWarp backend reaches 487,553 environment steps per second at 8,192 parallel environments, above mjlab's 391,556.

Source-provided article image: LocoMuJoCo: A Comprehensive Imitation Learning Benchmark for Whole-Body Control
arXiv

Interpretation

Whole-body imitation-learning and reinforcement-learning research has been split across single-robot codebases, manipulation-only suites and toy locomotion tasks, so changing the embodiment meant re-implementing data pipelines, retargeting methods, simulators and evaluation metrics. LocoMuJoCo puts that stack into one open-source library whose single API spans a serial NumPy/MuJoCo backend and parallel JAX/MJX and MjWarp backends, shipping 9 humanoids, 3 (musculo-)skeletal models and 4 quadrupeds. The first LocoMuJoCo version paired realistic humanoid models, motion-capture datasets and a common evaluation protocol but ran on CPU MuJoCo only; version 1.0 rebuilt the library around JAX, adding a GPU-parallel MJX backend, the retargeting pipeline and more embodiments, and the current version adds an optional MjWarp backend, further algorithms and the loco-mujoco-real deployment pipeline. The comparison with mjlab uses the same Unitree G1 model, with identical bodies, masses, actuators and 34 self-collision geometries, both on MuJoCo 3.11 at 50 Hz with four 5 ms physics substeps and the same solver settings; with full self-collision, MjWarp reaches 251,472, 442,715 and 487,553 environment steps per second at 1,024, 4,096 and 8,192 parallel environments, against mjlab's 135,086, 315,642 and 391,556.

Everything that varies across robots, tasks and research questions is a plug-in: observations, rewards, control functions, domain randomization, terrain, and initial and terminal state handling are separate, so switching from velocity tracking to motion imitation is a single argument change. Each plug-in is stateless and keeps its per-environment state in an explicit Flax dataclass passed through the step, so the whole environment step is JIT-compilable and one training loop serves every robot and both reinforcement and imitation learning. The retargeting pipeline builds on the SMPL-fitting approach of PHC and H2O, which targeted individual robots; here it becomes a configuration-driven pipeline in which adding a robot needs only a configuration file and 15 mimic sites in its model, and trajectories transfer between robots through an SMPL intermediate. Evaluation runs on the same 13 clips of the native LocoMuJoCo dataset across 11 embodiments with 5 seeds per cell, and trajectory similarity is measured with Euclidean distance, dynamic time warping and the discrete Fréchet distance over joint positions, joint velocities and the root-relative positions, orientations and linear velocities of the mimic sites.

Perspective

The demonstrated scope is a simulation benchmark plus two real-robot examples: evaluation over 11 embodiments and 13 native clips, and the deployment pipeline for the Unitree G1 and AgiBot X2 Ultra in loco-mujoco-real; the throughput measurements cover a single embodiment on a single GPU.

The throughput comparison measures a single embodiment on a single GPU, and mjlab's environment-level throughput includes its own heavier task logic; the real-robot evaluation switches off domain randomization and perturbations, so it measures tracking quality rather than robustness.