MimicX-Policies

MimicX-Policies

Paper (arXiv:2610.09055) | Hugging Face Paper Page | Project website

<!-- release-navigation-badges --> <div class="flex flex-wrap gap-2 items-center" style="display:flex;flex-wrap:wrap;gap:8px;align-items:center"> <a href="https://github.com/NEBULIS-Lab/MimicX"><img class="m-0 inline-block" style="margin:0;display:inline-block" src="badges/code.svg" alt="Code" height="32"></a> <a href="https://huggingface.co/Shuaijun/MimicX-Policies"><img class="m-0 inline-block" style="margin:0;display:inline-block" src="badges/policies.svg" alt="Policies" height="32"></a> <a href="https://huggingface.co/datasets/Shuaijun/MimicX-Assets"><img class="m-0 inline-block" style="margin:0;display:inline-block" src="badges/assets.svg" alt="Assets" height="32"></a> <a href="https://huggingface.co/Shuaijun/MimicX-Policies/blob/main/recommended/README.md"><img class="m-0 inline-block" style="margin:0;display:inline-block" src="badges/start.svg" alt="Start here" height="32"></a> </div> <!-- /release-navigation-badges -->

Pretrained G1 humanoid tracking policies for Dance, Football, Kung Fu and Tennis, plus additional video-driven motions.

Try a policy: choose a motion below or download it by task name. Reproduce the paper: use the original policies and multi-seed comparisons in the paper archive.

Current Demonstrations - October 14, 2026

Latest policies, matching assets and evaluation.

Parkour uses a separate MimicX + MPC recording, not a replacement PPO checkpoint. The original paper experiment archive remains unchanged.

Download by Task Name

After installing MimicX, run:

python -m pip install -e '.[artifacts]'
python -m mimicx.runtime.demonstrations download --task tennis --output artifacts

Choose tennis, football, dance, kungfu, track, stairs, forest, bench (Platform Jump), or descent. The downloader selects the matching checkpoint, reference and runtime files automatically. See the evaluation guide to run the policy.

Earlier scene policies remain in the October 11 release, including the locally retimed Parkour policy.

Paper Archive - Four Core Policies

Start here: per-task downloads and evaluation instructions. Download one policy and its matching reference; you do not need all 161 weights to try MimicX. Install the MimicX code and its tracking environment before execution.

These are the seed-101 final outputs from the paper experiments, with per-task settings and evaluation instructions.

Companion assets: Shuaijun/MimicX-Assets. Both repositories are available. The companion dataset contains the matching references, reproduction configurations and execution records.

Choose Your Level

Paper and Code

MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking

Paper: arXiv:2610.09055
Project website: https://nebulis-lab.com/MimicX
Code and installation: https://github.com/NEBULIS-Lab/MimicX
Paper source: https://github.com/NEBULIS-Lab/shuaijun-ICLR-paper-MimicX
Results and website source: https://github.com/NEBULIS-Lab/MimicX/tree/main/docs

Task-specific PPO checkpoints for MimicX humanoid motion tracking. The release includes Fixed Reference, Failure Curriculum, Task-Aware Refinement and MimicX.

Original Paper Archive Contents

Across both repositories: 161 unique checkpoint files, 94 recorded candidate-training endpoints, 10 bundled robot references and 12 complete core gate records. All methods and seeds are included. The 48 core trials use 44 distinct final checkpoints because some trials retain the same starting policy. Warmstarts, final policies and training candidates are listed separately in the catalog.

Download and Use

For complete reproduction, download this repository and its companion MimicX-Assets. Keep the two folders next to each other. See USAGE.md for preparing paper inputs, evaluating policies and replaying recorded acceptance decisions. prepare.py never starts training or GPU inference; it only validates files or prepares commands.

python prepare.py verify --root .

The command above verifies a complete snapshot. For a selective recommended download, follow its per-task instructions: prepare-eval checks the selected policy/reference pair without requiring unrelated archive files.

Policy and Evaluation Contract

Policies use G1 29-DoF tracking, PPO, observation normalization, task-specific warmstarts and registered robot references. Checkpoints retain actor, critic, optimizer and iteration state; do not discard observation normalizers. The core actors use 160 inputs and 29 outputs. Consult the checkpoint audit and archived configuration for each released model rather than assuming all additional cohorts share an identical network shape.

Core static methods use 250 continuation iterations; MimicX uses one search round with three 250-iteration candidates plus repeated verification. During the automated loop, the registered motion stays fixed. Dance and Kung Fu use pre-registered repaired references for MimicX; their component methods use the original references. Selected incumbents remain the final output when a candidate fails verification. A candidate-training endpoint is not necessarily an accepted policy. Scientific aggregate tables remain in the project website.

HLoop records use a fixed workload of eight policy rollouts, eight CPU diagnoses and one report selection, with zero PPO updates. They are distinct from the repeated candidate acceptance gate.

Data Sources and Reuse

Raw human videos, SMPL-family models/parameters, GVHMR and SONIC pretrained weights, and third-party motion datasets are not redistributed here. For the 14 licensed breadth references, the catalog identifies which source data to obtain separately. Supplementary policy weights and their execution records are included; those reference inputs are not claimed to be bundled. The historical lafan1_Form_1_stageii_g1 identifier belongs to AMASS/ACCAD Form 1 provided retargeting, not LaFAN1.

See THIRD_PARTY.md and reference_index.json in MimicX-Assets. The license covers the MimicX-authored files in this snapshot and does not grant rights to external datasets or dependencies. Robot reference arrays and recorded tracking errors do not contain SMPL model weights or source RGB frames.

Release Details

Checkpoints, motion references and experiment outcomes are preserved. Private machine paths are removed from public metadata. Telemetry is compressed losslessly, except that licensed breadth exports omit three reference-root coordinate columns while retaining all reported metrics.

The preparation tools check downloaded files automatically. Detailed file records remain available in the release index and source records. Use checkpoints from trusted sources; archived Python-tagged YAML files are records, not scripts to execute.

Browse the Files

POLICY_CATALOG.md provides a task-by-task download table.

Citation

@misc{liu2026mimicx,
  title={MimicX: Policy-in-the-Loop Supervision Refinement for Video-Driven Humanoid Motion Tracking},
  author={Shuaijun Liu and Chenglong Zhang and Xuhao Liu and Feiyang You and Yifan Liao and Shuyang Hao and Chaozhe Zhang and Chengyu Wu and Zhen Sun and Ningxin Su},
  year={2026},
  eprint={2610.09055},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.09055}
}