AMASS: Archive of Motion Capture as Surface Shapes
The largest unified human motion capture dataset. 15 mocap datasets fitted to SMPL body model. Foundation for humanoid motion generation.
Key Stats
| Metric | Value |
|---|---|
| Duration | 40+ hours of human motion |
| Source datasets | 15 (CMU MoCap, HumanEva, KIT, SFU, Eyes Japan, and more) |
| Body model | SMPL / SMPL+H / SMPL-X |
| Activities | Walking, running, sports, dance, daily activities, gestures |
| License | Custom academic (Max Planck Institute), requires registration |
| Origin | Max Planck Institute for Intelligent Systems |
What is AMASS?
AMASS solves a fundamental problem in human motion research: fragmentation. Before AMASS, the motion capture community had dozens of datasets in incompatible formats, marker sets, and skeleton definitions. AMASS unifies 15 of the most important optical mocap datasets by fitting all the raw marker data to the SMPL parametric body model, producing a consistent representation where every frame describes a full body shape and pose.
This unified representation makes AMASS the default training corpus for motion generation models (MDM, MotionDiffuse, MoMask), motion-language models (via HumanML3D, which adds text descriptions), and humanoid robot whole-body controllers. If you are training any model that needs to understand or generate human motion, AMASS is likely in your data pipeline.
Access
Note: AMASS requires registration at the MPI website and is restricted to academic/research use.
Related datasets
- HumanML3D -- AMASS + text descriptions for motion-language research
- CMU MoCap -- the largest single contributor to AMASS
- Unitree G1 Datasets -- real humanoid manipulation data
- NVIDIA GR00T -- humanoid teleoperation data