Success
A motion passes only if it reaches the final reference frame without any declared failure. All motions in an evaluation use the same policy and checkpoint.
BENCHMARK RESULTS
MotionChallenge Leaderboard
Last updated ·
| Rank | Model | Type | Organization | Mean per-body position error · MPBPE mmLower is better | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| No terminationEntire motion, including post-failure states | Early terminationUntil first failure, including the failure frame | ||||||||||
No models match your search.
Type: Dense = continuous dense policies (MLP, convolutional, graph or Transformer); FSQ = finite scalar quantization; MoE = mixture of experts; FB = forward-backward latent policy; Flow = flow matching. Sources are linked from each model card.
Each MuJoCo 3.17.0 batch evaluation reports success, completion and six tracking-error views. Only MuJoCo 3.17.0 batch results are published. Published evaluator ↗
Shared rules. Public code. Reproducible runs. Within each benchmark, robot and protocol, the core fixes data, physics, failures and scoring. Each entry identifies its checkpoint and inference implementation.
Rank follows the selected metric, always best first: higher success or completion, lower tracking error. The primary metric is No termination · XY only MPBPE (lower is better). Medals mark the top three. MPBPE averages distances over the benchmark’s declared body origins and frames, then weights motions equally. Early-termination errors cover the observed prefix. No-termination errors cover the entire motion.
Download results (JSON) ↓Completion
Mean fraction of motion completed before failure. Reported alongside tracking error to describe how much of each motion was completed.
Six tracking-error views
Each time range reports XY only, XY + yaw and no alignment. These remove horizontal translation, translation and heading, or neither. Height is retained in all six views.
Metric definitions →Citation
If you use MotionChallenge in your research, please cite:
@misc{motionchallenge2026,
author = {{MotionChallenge Team}},
title = {{MotionChallenge}: An Open, Reproducible Benchmark for Humanoid Motion Tracking},
year = {2026},
url = {https://motionchallenge-project.github.io/motionchallenge/}
}