MOVAS-Sync — Time-Aligned Multi-Modal Training Bundles
Video, audio, text and action streams delivered as hardware-synchronised bundles that drop directly into a training loop without alignment work.
Overview
Most of the cost of multi-modal training is not model code, it is alignment. MOVAS-Sync removes that: every stream in a bundle shares one hardware timebase, verified per bundle rather than assumed.
Bundles are shipped in WebDataset and MCAP so they stream directly into distributed training without a preprocessing pass.
Data Structure
Directory layout as delivered. Every clip is self-contained: no cross-referencing required to train on a single sample.
movas-sync/
├── shards/{shard_id}.tar # WebDataset
│ └── {sample}.{mp4,wav,json,npz}
├── mcap/{bundle_id}.mcap # ROS-compatible
└── manifest.parquetAnnotation Schema
Bundle
Stream inventory, duration, source scene.
Transcript
Diarised speech transcript with word-level timing.
Action
Control-stream events aligned to video timestamps.
Quality Gates
Published thresholds, enforced at ingest. Records that fail are rejected or flagged in the manifest — never silently included.
| Gate | Threshold | Method |
|---|---|---|
| Cross-stream drift | < 1 ms / min | Measured per bundle, recorded in the manifest. |
| Audio SNR | > 20 dB | Bundles below threshold are flagged, not silently included. |
| Transcript WER | < 8% | Human-verified on a 5% sample. |
Formats & Access
Export formats
Modalities included
Cloud delivery
Direct to your S3, GCS or OSS bucket. Manifest-driven incremental sync.
Air-gapped
Physical media transfer for regulated or offline environments.
Via Movas-OS
Stream and re-annotate in place through the Movas-OS platform.
Use Cases
Multi-modal pre-training
Joint video-audio-language representation learning.
Instruction grounding
Spoken instruction aligned to the action that followed it.
Audio-aware manipulation
Contact and material cues that vision alone misses.