Research & AI

Human gestures for AI research.

Gesture and expression data captured from Japanese performers. Tell us about the motions you need for training or evaluating AI models that work with human movement.

Inside the recording data

01

Record the performance in video and motion data

Reference footage preserves how the performance looks. FBX records the body motion, allowing the filmed gesture to be checked alongside its 3D movement.

02

Compare six interpretations of the same wave

Each wave pack contains six performances with differences in range, gaze, the pause after a gesture and head tilt.

  • T01A natural wave
  • T02A small, restrained wave
  • T03A big, energetic wave
  • T04Notice someone, then wave
  • T05Wave and wait for a response
  • T06Wave with a questioning head tilt

03

Keep the source and identify each performance interval

Wave packs keep the recording file (for example genki_Take002.fbx) as one file with its opening and closing T-poses. The table below lists the performance intervals used for the genki_Take002.fbx previews.

Preview intervals (frames) · 120 fps
IDStartEnd
T0110242104
T0224443528
T0338204880
T0452446344
T0566727728
T0680889164

Separate the action from the way it is performed

An energetic wave and a shy wave can differ in speed, range and whole-body movement. SOBURIMO aims to build data that lets researchers compare these differences in performance.

01

Generate motion with a specified personality or emotion

Capture conditions
Keep an action such as “greet someone with a right-handed wave” consistent while changing the direction: energetic, gentle or shy. Record the starting pose, position and how the action ends.
Preparing the data
Segment each action and pair it with its performance direction. Treat the intended direction separately from the impression reported by an observer.
Possible application
Explore training or evaluating models conditioned on both action and performance style, such as a gentle character greeting someone.

02

Generate or retrieve motion from text

Capture conditions
Specify the hand used, direction, travel and repetitions so the performed action can be identified clearly.
Preparing the data
Attach descriptions of the observed motion to time intervals, for example: “Raise the right hand to chest height and wave twice with small movements.” Annotation detail and Japanese or English coverage are agreed separately.
Possible application
Paired descriptions and skeletal sequences can support text-to-motion generation or natural-language motion retrieval, subject to model-specific preparation.

03

Compare generated motions under consistent conditions

Capture conditions
Repeat the same performance direction. When comparing performers, keep the instructions consistent and distinguish variation within a performer from variation between performers.
Preparing the data
Organise performer, session and take identifiers, then define training and evaluation splits. If impression ratings are needed, record the evaluators and criteria too.
Possible application
Evaluate whether a requested style is expressed and whether generated motions are varied. Perceived cuteness is not treated as a single objective ground truth.

04

Connect video observations with 3D movement

Capture conditions
The test capture includes three filmed reference views and skeletal data. Paired use requires checking time offsets and matching action intervals.
Preparing the data
Align video frames with skeletal timestamps. Review camera calibration, coordinate systems and occlusions as required. Fully synchronised, calibrated delivery has not been confirmed.
Possible application
Explore video-based 3D motion estimation or compare generated motion with a performance. Robotics additionally requires retargeting, joint-limit and contact checks, and control learning.

Captured files are not yet a task-ready training dataset

The verified inputs are test-capture FBX files and reference footage. Segmentation, captions, impression ratings, matched performer conditions and training/evaluation splits require preparation or review for the intended task. Supplying data does not guarantee a particular generated result or accuracy.

About the data

These specifications describe the current test recordings. Delivery formats and contents are agreed for each project.

Motion source
The test recordings currently under review are FBX at 120 fps with 73 bones, recording the motion of the body and finger skeleton over time. Before integration into research, bone names, hierarchy, coordinate system and units are reviewed.
Recording structure
Performances are recorded for each character. In the wave packs (5 packs), one FBX contains the 6 scheduled performances back to back: a natural wave, a small wave, a large wave, a wave after noticing someone, waiting for a response, and a head tilt. The performance packs (15 packs) contain 22–28 performances each, delivered as one cut file per performance.
Performance information
The performance sheet records the performance ID, action, emotion, intensity and direction notes. These are the directions given at the time of capture; they are not measurements of the performer's inner state, nor ground-truth emotion labels judged by AI.
Performance intervals
In the catalog, one performance runs from the moment the directed action begins until it ends. Each performance is mapped to the source file by its performance ID and start and end frames. The catalog currently offers 410 performances in 20 packs, but preparation of all recordings is not yet complete.
Calibration pose and delivery unit
The T-poses at the start and end of a recording are kept, and the motion source is handled as the complete recording file. Applying the motion to models with different proportions or skeleton structures requires adjusting the initial pose or retargeting. Please ask us about interval data or converted formats for research.
Reference video
1920×1080 at 60 fps, with three filmed viewpoints and an avatar screen reference. Video delivery requires a separate review of usage permissions.
Video alignment
The video is reference material for checking the actual performance. The current previews align the video and the FBX using an estimated time offset, so this data does not guarantee exact frame synchronization. If you plan to use pairs of video and 3D motion for training or evaluation, we agree the required synchronization accuracy and the scope of re-checking in advance.
Delivery requirements
Actions, data volume, file formats, skeleton structure, labels and reference video. Retargeting and any additional processing are discussed in advance.

Motion data describes body movement; it is not a measurement of robot joint torque or contact forces. Please confirm that the available information fits your intended use.

Discuss research and AI use

Contact us for pricing

Quotes are based on the motions and volume of data needed, along with the scope of research or commercial use. You can enquire about a subset of the data.

Evaluation data can also be discussed after reviewing your intended use and the conditions for providing it.

Usage terms and performer consent

AI training, sharing within or outside your organisation, and commercial use are addressed in the agreement. Only data covered by the relevant performer permissions can be provided.

The standard license does not permit AI training. Research and AI use require separate review and agreement on terms.

How to get started

  1. 01

    Tell us about your project

    Share your research or development goals, the motions you need and your timeline. You can contact us before your specification is final.

  2. 02

    Review data and usage

    We discuss the data, formats and sharing requirements, then check performer permissions to determine what can be provided.

  3. 03

    Quote and agreement

    Pricing reflects data volume and usage terms. We agree on the contents, price and delivery schedule before proceeding with a contract.

Research examples for these applications

Contact us

Tell us your research or AI use case and the data you need. Terms and pricing are discussed individually.