Behavioral Training Data

The training data that gives robots emotional intelligence.

Physics-validated, semantically-tagged group interaction datasets — purpose-built for humanoid robotics and game AI teams who need real human behavior at scale.

Request dataset access See what's inside

Consent-documented • IRB-reviewed • Physics-validated • Delivery in <48h

4.2M+
Annotated interaction frames
98.6%
Emotion-classification accuracy
312
Unique group scenarios captured
<48h
Typical dataset delivery SLA

Purpose-built for teams in

Humanoid Robotics TeamsAAA Game StudiosDefense & SimulationAcademic AI Research
The Data Gap

The existing datasets weren't designed for this problem.

Humanoid robotics and game AI have outpaced the data that trained them. The bottleneck isn't compute — it's the absence of authentic, structured human behavioral data.

SYN-P01

Single-face datasets fail in the real world.

Most emotion AI tools train on one isolated face. Robots and NPCs operate in crowds, meetings, and social environments — and the existing data simply doesn't reflect that reality.

SYN-P02

Annotation platforms give you labels, not physics.

Generic labeling tools tag faces and joints. They don't encode the physical dynamics — velocity, weight transfer, proxemics — that make simulated motion convincing and robotics safe.

SYN-P03

Consent and provenance stop production cold.

Scraping social video creates legal liability. Legal review of dataset provenance slows every deployment. Teams either ship shaky data or waste months sourcing clean alternatives.

The Synaphex Stack

Four layers of structure that no other dataset has.

SYN-S01

Group Behavioral Capture

Multi-person interaction sessions — natural conversations, physical tasks, emotional scenarios — recorded with full participant consent across controlled and naturalistic settings.

SYN-S02

Physics-Validated Annotations

Every skeleton frame is run through physics simulation validation. Force vectors, contact states, and biomechanical plausibility checks eliminate the synthetic jitter that breaks robot policies.

SYN-S03

Semantic Behavior Layer

Proprietary tagging system encodes social intent, emotion valence, conversational role, and relational dynamics — making datasets queryable by behavior, not just by anatomy.

SYN-S04

Consent-Clear Provenance

Every participant signs informed consent. Dataset cards document collection conditions, demographic coverage, and IRB status — so your legal team signs off in days, not months.

Dataset Architecture

Engineered for the way ML teams actually work.

Every structural decision in the dataset — from annotation schema to delivery format — was designed with the training pipeline in mind, not the annotation team.

SYN-F01

Facial Action Unit Streams

AU-coded facial data synchronized frame-by-frame with body keypoints, gaze vectors, and vocal prosody markers. 72-point facial mesh at 120fps.

SYN-F02

Multi-Modal Fusion

RGB, depth, and inertial data merged into a single unified timeline. No synchronization work on your end — the pipeline handles it.

SYN-F03

Scenario Taxonomy API

Query datasets by scenario type, emotional context, group size, or interaction phase. Pull exactly the slice your training run needs.

SYN-F04

Physics Engine Integration

Pre-validated for MuJoCo, Isaac Sim, and Unreal Physics. Import directly — every motion sequence is physically plausible out of the box.

SYN-F05

Custom Capture Programs

Don't see the scenario you need in our catalog? We design bespoke capture protocols for your specific robot embodiment or game character archetype.

SYN-F06

Benchmark Evaluation Sets

Held-out evaluation splits with gold-standard annotations for emotion, gaze, pose, and intent — so you can track model progress against a fixed reference.

Applications

Two industries. One missing ingredient.

Humanoid Robotics

Train robot social navigation, handover behaviors, and human-robot interaction policies on real multi-person dynamics — not synthetic mocap.

Figure AI 1X Technologies Agility Robotics

Game AI & NPCs

Give NPCs authentic emotional reactions, group crowd behaviors, and physically-believable motion that breaks uncanny valley for good.

EA Sports Ubisoft Riot Games
"The only vertically-integrated source of multi-person behavioral data engineered for the physical AI revolution."
Get Access

Stop waiting for data that doesn't exist yet.

Synaphex datasets are available under research and commercial licenses. Tell us what you're building and we'll identify the right dataset or design a custom capture program.

Contact the data team

Research licenses available • NDA-protected custom programs • Dataset previews on request