FLAGSHIP DATASET
The Data Engine for Physical AI

Data moves robots.
MOVAS moves data.

HUMAN WORKER · HEAD-MOUNTED EIS OFF CONTACT MOVAS-OS EGO VIDEO HAND POSE · MANO CONTACT STATE OBJECT 6-DoF RETARGET · IK SIM VALIDATED ✓ HUMANOID · RETARGETED POLICY REAL WORK → ROBOT-EXECUTABLE BEHAVIOR
For teams shipping HumanoidsBimanual manipulationWorld models
100%Employer-signed
49,000Trainable hours
6Layers per clip
5 daysTo an evaluation subset
MV-FEED-003Real world feed

Inside three clips from the dataset.

More samples
MV-EGO-100KMOVAS-Ego 100K · flagship dataset

MOVAS-Ego 100K

Scene Distribution

31,428 HRS

from the tasks robots struggle with most.

Contact-richBimanualTool-use
Commercial Operations
25,636 h  25.6%
Food Preparation
17,770 h  17.8%
Domestic & Care Services
17,200 h  17.2%
Industrial & Production
11,000 h  11%
General Retail
9,823 h  9.8%
F&B Storefront
7,466 h  7.5%
Specialized Services
5,397 h  5.4%
Handcraft & Fine Assembly
3,461 h  3.5%
Market, Agri & Mobility
2,247 h  2.2%
PREMISES 80.6%
RESIDENTIAL 17.2%

Residential capture is dispatched staff, not consumers.

All nine scene families
CONTACT DENSITY × VOLUME contact-rich band 0 25 50 75 100 hours in inventory → contact density → CommercialFoodDomesticAssemblySortingGeneralF&BSpecializedHandcraftMarket, biggest, least contact
Bubble area ≈ hours Relative scale, not a measured share Industrial shown as its two published sub-classes; the four points in the band sum to 31,428 h
All nine families
LAYER ALIGNMENT

Every layer, same moment.

WHAT SHIPS IN EVERY CLIP

What ships inside a single clip.

COMPONENTS  0 / 9
CAPTURED · EVERY CLIPRecorded at the sensor. On 100% of clips.
01

Stereo egocentric video

02

Inertial stream

03

Calibration

04

Frame integrity

05

Task annotation

ENGINEERED · MOVAS-OS OUTPUTRefined by Movas-OS. Coverage follows the published yield.
06

Hand pose

07

Contact state

08

Object pose

09

Retargeted trajectory

Five captured. Four engineered. One take. Full specification →

LAYER ALIGNMENT ON A SINGLE CLIP
clip_00841.mp4SYNC DRIFT 0.4 ms / 60 s
SOURCE · RGB
L1–L5 CONTROL
4D GROUND TRUTH
MOCAP
3D ASSETS
RETARGET · 5 EE

Hardware-synchronized. UMI · LeRobot · RDT · OXE · MOVAS native.

MV-ACT-004Action coverage

Four contact profiles, not one average.

Assembled by difficulty, not by scene.

01

Intermittent

Brief contact, rigid objects

Shelf replenishment · counter service
02

Sustained

Hands stay on while the shape changes

Dough work · folding · plating
03

Tool-mediated

Force through a held tool

Torque check · appliance service
04

Fine

Sub-millimeter clearance

Tweezer insertion · connector seating

Assigned per clip at L5, not inferred from the scene family.

MV-OS-006Movas-OS · the engine

From pixels to actions.

VLA models

vision → language → action

World models

geometry → dynamics → prediction

Robot learning

demonstration → retargeting → policy

Spatial intelligence

depth → 3D → 4D
How MOVAS supports embodied intelligence
REFINEMENT PIPELINE

We engineer it, not just collect it.

HAND · 0.936DoF · LOCKCONTACT · 386F
01FILTER66K h
024D RECON<3 px
03HAND0.93
046DoFLOCK
05RETARGET5 EE
06VALIDATE87/100
movas-os · ingest streaming
ingest → egocentric/clip_00841.mp4 sync → imu 476 Hz drift 0.4 ms PASS hand → HaMeR conf 0.93 L/R PASS object → 6-DoF mug_blue_001 PASS contact → frames 218-604 bimanual PASS retarget → shadow_hand ik solved PASS sim → isaac replay PASS score → 87 / 100 PRO TIER ────────────────────────────────────────────── ingest → egocentric/clip_00842.mp4 sync → imu 476 Hz drift 0.6 ms PASS hand → HaMeR conf 0.71 L only LOW fallback → WiLoR conf 0.88 PASS object → 6-DoF pan_steel_014 PASS contact → frames 91-377 tool-mediated PASS retarget → parallel_2f aperture 62 mm REJECT retarget → shadow_hand ik solved PASS sim → isaac replay PASS score → 79 / 100 STD TIER ────────────────────────────────────────────── ingest → egocentric/clip_00843.mp4 sync → imu 476 Hz drift 1.4 ms FAIL reject → gate: drift < 1 ms / 60 s → written to rejection set ────────────────────────────────────────────── ingest → egocentric/clip_00844.mp4 sync → imu 476 Hz drift 0.3 ms PASS hand → HaMeR conf 0.96 L/R PASS
6containerised stages
34%rejected at gate 01
49%reach delivery
100%version-pinned

Rejections ship on the data card too.

OPEN EVALUATION

MEVB

A yardstick for embodied policies, not video understanding. Open tasks, sealed splits, independently reproduced.

Open core · reproducible · academically governed
MV-FAQ-006Common questions

Before you ask.

FAQ

Evaluation subset of up to 50 hours in five business days. A full scene family at Control L1–L5 in twelve, seventeen with aligned layers. The hours exist already; nothing here is a collection schedule.

Name the scene family, sub-class and the failure mode your policy is stuck on. Work that is not in inventory can be re-captured at a contracted site, priced per site-month.

Control Control L1–L2 on the full corpus; L3–L5 follow the published yield. Optional aligned layers add 4D, motion capture, 3D assets, retargeting and the feasibility record — all time-locked to the same clip.

Yes. MOVAS signs with the operating business first and the individual second, and every clip carries a consent id traceable to both. An individual cannot license their employer’s premises, processes or the colleagues in frame.

Faces, plates, screens and documents are redacted on-device before upload; precise GPS is dropped at the edge; pre-redaction footage is destroyed on a 90-day schedule. Cross-border delivery follows each jurisdiction’s own data-transfer procedure, scoped before a program starts. Air-gapped delivery available.

Your models need the physical world.

Request a Subset

Data moves robots. MOVAS moves data.