LIVE CAPTURE

Group emotion data
robots can actually
learn from.

Sentic captures multi-person emotional dynamics, layers semantic intent, and validates every motion against physics — so your AI trains on truth, not noise.

4.2M+
labeled frames
18
emotion classes
3–9
people per scene
99.1%
physics-verified

Existing emotion datasets fail
under real-world conditions.

01

Single-actor footage ignores how emotions shift in group dynamics — your robot reads a crowd of two as a crowd of one.

02

Labels stop at facial action codes. There is no semantic layer connecting expression to conversational intent or social context.

03

Motions lack physics constraints — models trained on bad kinematic data hallucinate impossible body positions at inference time.

Three layers of truth,
not one layer of labels.

Sentic captures real group interactions simultaneously across face, body, and proxemics — then enriches every frame with a semantic intent layer and validates the full sequence against physics engine constraints before any data leaves our pipeline.

Group capture → Semantic enrichment → Physics validation

Three systems. One dataset you can trust.

Pillar 01 — Group Capture

Simultaneous multi-person emotion capture at scene level.

Existing datasets isolate individuals. Sentic captures every person in a scene simultaneously — tracking face landmarks, full-body joint kinematics, gaze vectors, and proxemic distances between all participants in real time.

Group dynamics change everything. A raised eyebrow means one thing in isolation; it means something entirely different when three people are watching. Our data captures that difference.

136
Face landmarks / person

MediaPipe + custom AU classifier trained on in-house annotator consensus.

33
Body joints tracked

Full-body pose in 3D world coordinates, not 2D projection artifacts.

9
Max simultaneous actors

Crowd-level social scenes captured with per-person identity tracking.

120fps
Capture frame rate

High-speed capture preserves micro-expression onset and offset timing.

// semantic_layer.schema
emotion_class 18 discrete classes + continuous valence/arousal values per 250ms window
intent_vector Social intent annotation: agreement, challenge, submission, humor signal, 14 types total
context_tag Scene-level context: setting type, power dynamics, relationship proximity, interaction phase
sync_score Inter-person emotional synchrony metric — unique to Sentic, absent from all other datasets
Pillar 02 — Semantic Layer

Every frame knows what the emotion means, not just what it looks like.

FACS codes and AU labels tell you a muscle moved. They cannot tell you whether that micro-expression signals amusement, social compliance, or suppressed anxiety in a given group context.

Sentic appends a semantic intent vector to every annotation — linking expression data to conversational and social function so your model learns the why alongside the what.

Pillar 03 — Physics Validation

Biomechanically valid or it does not ship.

Every motion sequence in the Sentic pipeline passes through a physics engine validation step. Joint angles are checked against anatomical limits. Velocity and acceleration profiles are verified for biomechanical plausibility. Sequences that fail are flagged and re-captured, not patched.

The result: models trained on Sentic data do not hallucinate impossible poses at inference time — a critical requirement for any robot operating in proximity to humans.

Joint angle validation

Elbow, shoulder, spine, and hip angles verified against published human ROM limits across 7 joint types.

Velocity & acceleration profiling

Frame-to-frame kinematic derivatives checked for biologically plausible movement speeds — flags capture artifacts and jitter.

Re-capture, not patching

Sequences failing validation are discarded and re-recorded. We do not interpolate or synthetically patch failed frames.

Built for teams that cannot afford to train on noise.

01 — Humanoid Robotics

Give your robot the social awareness to operate near humans safely.

Humanoid robots entering real workplaces and homes need to read group emotional states — not just individual faces. Sentic datasets train the models that power socially-aware navigation, interaction timing, and proximity management.

Series B+ humanoid teams
Collaborative robot researchers
Physical AI data teams
02 — Game AI Studios

NPCs that react to the whole room, not just the closest player.

Game AI characters trained on Sentic data understand crowd emotional dynamics — how tension builds in a group, how laughter propagates, when a crowd turns. The result is believable social AI that players feel rather than analyze.

NPC behavior engineers
Motion capture animators
Crowd simulation researchers
03 — Affective Computing APIs

A foundation dataset your emotion recognition models can grow from.

API companies building emotion recognition at scale need diverse, densely-annotated training data with strong social-context coverage. Sentic licenses full annotation schemas and delivery pipelines for direct integration into your training infrastructure.

Affective computing platforms
Emotion AI API providers
Social AI research labs

What you receive with a Sentic dataset license

Feature
Typical Datasets
Scale AI / Appen
Sentic
Group scenes (3+ people)
Rare
Occasional
100%
Semantic intent labels
None
Basic
18 intent types
Physics engine validation
None
None
Every frame
Emotion synchrony metric
None
None
Unique to Sentic
Body + face + gaze combined
Separate streams
Partial
Fused, aligned
Re-capture policy on fail
Patched
Patched
Full re-capture

Your AI models deserve data that holds up in the real world.

Sentic works with robotics teams, game AI studios, and affective computing companies on a direct license basis. Tell us about your training data requirements and we will respond within one business day.

NDA-protected previews available. Enterprise SLAs on request.

// what to expect
01
Response within 24 hours

A Sentic data engineer reviews your use case and sends a scoped proposal.

02
NDA-protected sample preview

Review 500 annotated frames with full schema documentation before any commitment.

03
Custom dataset scoping

License existing data or commission a custom capture batch tuned to your model requirements.