Can Movas-OS run inside our own VPC?

Yes. Movas-OS deploys as a managed service, inside your VPC, or fully air-gapped. In the VPC and air-gapped configurations your data never leaves infrastructure you control.

How much annotation time does the AI-assisted workflow actually save?

Up to 70% versus a standard manual pipeline on comparable footage. The saving comes from auto-segmentation, multi-camera track propagation and LLM-assisted pre-labelling — humans review and correct rather than label from scratch.

Does Movas-OS work with data we collected ourselves?

Yes. Ingest accepts standard video containers, ROS bags and MCAP. The ontology and annotation layers are agnostic to where the footage came from.

What does "sovereign provenance" mean in practice?

Every asset carries an immutable audit trail of who touched it, what transformation was applied and under what licence it entered the system. Combined with automated PII scrubbing and cryptographic watermarking, this is what makes the output defensible in a commercial training context.

SaaS Infrastructure

Movas-OS — The Engine Behind Every MOVAS Dataset

An industrial-grade end-to-end platform for ingestion, high-density annotation, curation and alignment evaluation — designed for complex multimodal assets, and available for your own footage.

Request a Walkthrough

[ingest] clip_00841.mp4 ok [sync] imu 200Hz drift 0.4ms ok [hand] HaMeR L 0.94 R 0.91 ok [object] 6-DoF mug_blue_001 ok [contact] frames 218-604 ok [retarget] shadow_hand ik solved ok [retarget] franka_grip ik solved ok [sim] isaac replay PASS [score] 87 / 100 -> PRO TIER ────────────────────────────────── [ingest] clip_00842.mp4 ok [hand] HaMeR conf 0.71 L only [fallback] WiLoR conf 0.88 ok

Platform Modules

AI-Assisted Workflow

LLM-augmented preprocessing with auto-segmentation and multi-camera tracking. Cuts manual annotation time by up to 70% versus standard pipelines.

Ontology Management

Build nested schemas that map visual features directly to physical properties. Supports dynamic property hierarchies and cross-task taxonomies.

Sovereign Provenance

Automated PII scrubbing, cryptographic watermarking and full IP audit trails. Air-gapped deployment available; ISO 27001 aligned.

Egocentric Processing

Native hand-pose extraction, 4D scene reconstruction and human-to-robot retargeting across five end-effector families.

Simulation Validation

Replay retargeted trajectories in Isaac Sim or MuJoCo to reject physically implausible sequences before they enter training.

Format Export

One-click export to UMI, LeRobot, RDT and Open-X-Embodiment without re-processing.

Processing Pipeline

Six stages from raw footage to trainable trajectories. Each stage is independently containerised and version-pinned, so a result can always be traced back to the exact model that produced it.

raw footage
    ↓
[1] candidate filtering      view classifier + VLM task filter + hand visibility gate
    ↓
[2] 4D reconstruction        metric depth + camera trajectory + scene geometry
    ↓
[3] hand pose extraction     MANO parameters, bimanual state, temporal smoothing
    ↓
[4] object 6-DoF tracking    open-vocab detection + video segmentation + pose
    ↓
[5] human-to-robot retarget  wrist mapping + IK + contact-preserving optimisation
    ↓
[6] simulation validation    physics replay, plausibility gate, quality scoring
    ↓
UMI / LeRobot / RDT / Open-X-Embodiment / MOVAS native

Deployment Options

Managed Cloud

Fastest to start. MOVAS operates the infrastructure; you get a workspace and an API key.

Your VPC

Runs inside your cloud account. Data never crosses your network boundary.

Air-Gapped

Fully offline installation for regulated environments. Updates delivered on physical media.