Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('test'): ('csv', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

SuperMimicLite

G1 motion tracking with MimicLite and Uni-SuperMimic. Includes PPO and SuperMimic training, pretrained policies, ONNX export, and a fixed 228-motion evaluation.

Code · Models · Training · Evaluation

This hub hosts 3,692 G1 training motion files and 228 evaluation motions, mirrored byte-for-byte from the pinned public Any4 HDMI sources. Full selection lists, hashes, source revisions, and original source terms are retained.

Included motion files

Data NPZ files Download
LaFAN40 40 Files
100STYLE 3,573 Files
Ground motions 43 Files
AMASS-hard training view 15 Files
Extreme motions 5 Files
Extreme demo, non-SONIC 3 Files
Real-pico 13 Files

Total motion payload: 364 MB. Every hosted motion matches the files used by the released training recipes; no new retargeting or resampling is applied. The dataset directories retain attribution and source/license references.

hf download han-xudong/SuperMimicLite --repo-type dataset \
  --include "training/public/*" --local-dir data

Not hosted: SONIC motions, the twelve SONIC-derived extreme-demo clips, and both high-dynamic sequences containing SONIC segments. The complete Mixed recipes still require these motions under their original source terms. The extreme-demo directory intentionally contains only three of the fifteen motions described by its unchanged canonical upstream manifest.

Evaluation motions

Subset NPZ files Download
Locomotion 80 Files
Manipulation 48 Files
Ground 60 Files
Dance 40 Files

The complete fixed evaluation set is 228 NPZ files, 135 MB. Original 120 Hz G1 trajectories, the frozen motion order, selection manifests, source attribution, and ChingMu access terms are unchanged. Policy evaluation runs at 50 Hz. The four direct G1 source revisions are recorded in evaluation/sources.json.

uv run scripts/fetch.py data

This downloads all four subsets to data/motiondecode/ from this repository at a pinned revision and checks file hashes and the frozen evaluation order.

Collections

Collection Policies Motion files Control frames at 50 Hz
LaFAN40 P-L, S-L 40 441,140
Mixed P-M, S-M 131,373 unique 50,712,626 unique
Evaluation All four 228 —

Mixed preserves ten source groups and their original sampling weights. Counting overlapping memberships gives 131,897 entries and 51,098,520 frames. The fixed evaluation includes Locomotion80, Manipulation48, Ground60 and Dance40.

Source revisions

Training source Upstream repository
any4hdmi-g1-lafan elijahgalahad/any4hdmi-g1-lafan
any4hdmi-g1-100style elijahgalahad/any4hdmi-g1-100style
any4hdmi-g1-ground-motions elijahgalahad/any4hdmi-g1-ground-motions
any4hdmi-g1-amass-hard elijahgalahad/any4hdmi-g1-amass-hard
any4hdmi-g1-extreme-motions elijahgalahad/any4hdmi-g1-extreme-motions
any4hdmi-g1-extreme-demo elijahgalahad/any4hdmi-g1-extreme-demo
any4hdmi-g1-high-dynamic-sequences elijahgalahad/any4hdmi-g1-high-dynamic-sequences
any4hdmi-g1-real-pico elijahgalahad/any4hdmi-g1-real-pico
BONES-SEED / SONIC Official gated source

The four evaluation source repositories and immutable revisions are listed in evaluation/sources.json. The card's source_datasets YAML records the direct mirrored datasets and original sources. BONES-SEED is included for full-training selection and metadata provenance; no SONIC motion payloads are hosted here.

Metadata files

  • sources.json: hosted dataset paths, pinned upstream references, and metadata checksums.
  • training-admission.json.gz: selected filenames, file sizes, hashes and frame counts; no motion coordinates or annotations.
  • metadata/: the canonical SONIC format manifest and exact selection filters used by the original recipes.

Prepare

git clone https://github.com/han-xudong/SuperMimicLite.git
cd SuperMimicLite
uv sync --locked
uv run scripts/bootstrap.py

# The complete 228-motion evaluation set from this project's pinned data release.
uv run scripts/fetch.py data

# Complete LaFAN40 from this project's pinned data release.
uv run scripts/training_data.py prepare --root data/training --data L

For Mixed, accept the BONES-SEED license, request access through its official repository, download g1.tar.gz, and extract g1/csv. Then prepare the locked conversion environment:

uv run scripts/bootstrap.py --training
uv sync --project environments/train --locked
uv run scripts/fetch.py asset --training-asset
uv run scripts/training_data.py prepare --root data/training --data M --sonic-csv /path/to/g1/csv --workers 8

The preparation script stages CSV conversion separately, installs only admitted motions, and verifies hashes and selection counts. Existing matching files are reused. Training recipes, budgets and source weights remain as documented in the training guide.

Terms and attribution

Training data includes Motion Data by Bones Studio. Use of the underlying dataset is subject to the BONES Motion Capture Dataset License Agreement. BONES-SEED prohibits public mirrors and redistribution of its motion files. All mirrored trajectories retain their original source terms, including LAFAN1's CC BY-NC-ND 4.0 and AMASS's research and redistribution terms. Public source availability is not an additional license; this mirror does not grant new rights to the underlying data. Training/evaluation collections may overlap.

SuperMimicLite's original code is Apache-2.0. See THIRD_PARTY.md for source attribution.

Downloads last month
42

Models trained or fine-tuned on han-xudong/SuperMimicLite