Embodied AI data foundry

Data for robots,
built to specification.

Real dataset sampleSemantic + pose overlay
CAPTURE29.67 fps
TRACKING24 body · 52 hand joints
ANNOTATION32 segments · 11 actions

20-second web preview · Full 59.78-second record and QA package below

License action-labeled datasets or commission a custom collection program. Synergy delivers synchronized, quality-controlled data for VLA models, manipulation policies, and embodied world models.

Bimanual actionsMulti-view RGB6-DoF poseGripper state

Built for robot learning teams

VLA FINE-TUNINGIMITATION LEARNINGOFFLINE RLTASK PLANNING
Data assets

Purpose-built data for
distinct learning problems.

Not all embodied data trains the same capability. We package each asset around the model behavior it can credibly support.

01 / EGO + UMIPolicy learning
EPISODE SCHEMASYNCHRONIZED
VisionGlobal + dual wrist RGB30 fps
TrajectoryJoint + control valuesfull
Pose6-DoF end effectoraligned
GripperLeft + right aperturestate

Bimanual manipulation demonstrations

Proxy-tool human demonstrations across household, light-industrial, and outdoor tasks, with full action trajectories and segment-level annotations.

Parquet episodesMulti-viewInspection records
02 / EGOCENTRIC POSEMotion + semantics
SAMPLE CAPTURECALIBRATED
Video4096 × 1536 stereo29.67 fps
Body24 tracked joints~90 Hz
Hands26 joints per handarticulated
Labels32 temporal segments11 actions

Egocentric human activity and pose

First-person activity capture with calibrated stereo video, whole-body and articulated hand tracking, plus frame-aligned English action descriptions.

Head + body + handsStereo calibrationAutomated QA
03 / MULTI-SCENE RGB-DPlanning + world models
CAPTURE SYSTEMMULTI-VIEW
VISION3 calibrated RGB-D cameras
INFRARED4 synchronized channels
LABELStask phase · causality · exceptions
SCENES7 workplace scenario families

Multi-scene task-planning corpus

Long-horizon human activity for task planning, scene understanding, video-language pretraining, and metric 3D reconstruction—not low-level robot control.

RGB-D + IRTask phasesCausal labels

Representative specifications from sample assets. Final availability, volume, quality metrics, and license rights are confirmed during technical diligence.

Inside a second sample delivery

Different task.
Same evidence standard.

This 59.91-second garment-handling record spans six clean bimanual action segments—buttoning, smoothing, folding, placing, lifting, and unfolding—with calibrated video, full hand coverage, and machine-readable QA.

Request the sample manifest
Real sample · bimanual garment handling59.91 sec
VIDEO1280 × 760 web preview
CAPTURE30 fps · 59.91 sec
TRACKING5,400 frames · 100% hand coverage
LABELS6 actions · 0 noise intervals
Production system

The pipeline is
the product.

Standard operating procedures, versioned annotations, inspection records, and coverage design turn collection capacity into repeatable data production.

Sample recordConference-room object transfer
5,402tracking frames
32temporal segments
11atomic action types
45.67sclean clip · 100% hand coverage
openliftwalkplacealign
Sample inspectionMachine-readable report
12automated
checks run
Camera calibrationVALID
Video / tracking sync1.001
Motion anomaliesNONE
Semantic noise1 EXCLUDED
Sample manifestAuditable by design
meta.jsonCapture manifest
camera_paramsIntrinsics + extrinsics
trackingDataFrame-level pose
segments_ENAction semantics
quality_inspectionAutomated QA

Each component is independently inspectable and linked by timestamps.

Custom data programs

Your task roadmap becomes
our collection plan.

Define the tasks, objects, environments, capture format, and quality bar. We deliver QC-passed data against agreed acceptance criteria.

Scope a collection pilot
01

Scope the target behavior

Translate model gaps into tasks, variations, sensors, and measurable acceptance criteria.

SPECIFY
02

Calibrate with a pilot

Validate the capture method, schema, annotation, and QC loop on a bounded first delivery.

PILOT
03

Scale and refresh

Run parallel collection streams and deliver new, QC-passed episodes against your roadmap.

OPERATE

Collection methods

Robot teleoperationUMI / proxy toolEgocentric demonstrationMotion captureWearable sensorsSimulation
Start with the model gap

Tell us what your robot
needs to learn next.

We will map the behavior to a dataset, evaluation slice, or custom collection program.

Request a technical consultation