MimicX-Assets

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

Reference motions and reproducibility assets for MimicX. Start with a task's robot reference and its matching policy, preview the recorded executions, or inspect the complete experiment archive. Both repositories are available.

Current Demonstrations - October 14, 2026

Latest policies, matching assets and evaluation.

Motion	Recommended checkpoint	Reference and video	Execution
Tennis Swing	Download	Files and video	518 steps / 10.36 s
Football Juggling	Download	Files and video	454 steps / 9.08 s
Dance Sequence	Download	Files and video	783 steps / 15.66 s
Kung Fu	Download	Files and video	1061 steps / 21.22 s
Track Running	Download	Files and video	97 steps / 1.94 s
Stair Ascent	Download	Files and video	449 steps / 8.98 s
Forest Traversal	Download	Files and video	404 steps / 8.08 s
Platform Jump	Download	Files and video	58 steps / 1.16 s
Stair Descent	Download	Files and video	240 steps / 4.80 s

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 Tasks

These four references match the recommended final MimicX policies. Download only the pair you need; the full telemetry archive is not required to run it.

Motion	Robot reference	Matching policy	Instructions
Dance	Download reference	Final policy	Run Dance
Football	Download reference	Final policy	Run Football
Kung Fu	Download reference	Final policy	Run Kung Fu
Tennis	Download reference	Final policy	Run Tennis

Recommended reference manifests retain each exact policy/reference pairing and configuration. The reference is a robot-motion NPZ, not the source human video. Install the MimicX code and tracking environment before execution.

Watch Original Paper-Archive Policies

Eight existing robot-only videos, organized by task and method. Each link opens the HF file viewer, where the original MP4 can also be downloaded. Videos show individual recorded rollouts; numerical comparisons use the registered repeated evaluation protocol. The Fixed Reference and MimicX columns remain separate.

Motion	Fixed Reference	MimicX
Dance	Watch / download	Watch / download
Football	Watch / download	Watch / download
Kung Fu	Watch / download	Watch / download
Tennis	Watch / download	Watch / download

Choose Your Download

Goal	Entry point	What you need
Try a policy	Current demonstration assets	One robot reference and its paired policy
Reproduce the core study	Usage guide, registered input map	Both snapshots and the pinned runtime
Inspect execution-gate decisions	Gate index, replay audit	Recorded telemetry and the code's gate implementation
Inspect HLoop execution	Workload, summary	Archived workload, executor and timing records
Browse all references	Reference catalog	Bundled motions and separately licensed inputs

Reproduce and Verify

For a single task, follow its Instructions link above. The supplied prepare-eval command checks the policy/reference pair and prepares execution commands without launching them.

For the complete archive, download both repositories into sibling directories:

hf download Shuaijun/MimicX-Policies --local-dir MimicX-Policies
hf download Shuaijun/MimicX-Assets --repo-type dataset --local-dir MimicX-Assets
python MimicX-Assets/prepare.py verify --root MimicX-Assets
python MimicX-Policies/prepare.py verify --root MimicX-Policies

Full-snapshot verification requires all indexed files; it is not the check for an intentionally selective download. See USAGE.md for reconstructing the 22-file core input bundle, preparing evaluations and replaying all 12 core gates.

Original Paper Archive

Artifact	Scope
Robot reference motions	10 bundled NPZs; 14 separately licensed references
Core experiment records	48 final trials across four tasks, four methods and three seeds
Additional video records	24 final trials
Provided-motion breadth	24 final trials
Fresh AMASS/ACCAD retargeting	4 final trials
Execution-gate evidence	12 complete core gates, with all accept/protect outcomes retained
HLoop evidence	Fixed workload, five measured repetitions and warmup records
Robot-only videos	8 recordings: four tasks, two methods

The companion Policies repository holds 161 unique checkpoints and indexes 94 candidate endpoints. Trial identities, candidate endpoints and distinct network files are different counts. Scientific summary tables live in the project website.

During the automated loop, the registered motion stays fixed. The recommended Dance and Kung Fu references are the registered repaired motions; their fixed-reference comparisons use the original motions. HLoop's workload uses the registered Fixed Reference policies, eight rollouts, eight diagnoses and one selection, with no PPO updates. Full policy/training details are in the Policies model card.

Paper and Code

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

• Paper: arXiv:2610.09055
• Project website
• Code and installation
• Paper source
• Project website source and numerical downloads

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

REFERENCE_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}
}