VLA models, world-action models and world models each demand different data. We capture first-person and multi-camera stereo data in real scenes, annotated with action primitives, task instructions and state changes — aligned directly with model training needs.
From VLA to world models, each paradigm needs its own data shape — so we organize capture and annotation accordingly.
Taking vision and language instructions as input and directly outputting action sequences, VLA training requires strictly aligned vision-language-action trajectories: frame sequences, action primitives, natural-language instructions and temporal segmentation.
A new paradigm jointly modeling future video and future actions. Dyna-2, pre-trained on over one million hours of first-person human manipulation video, demonstrated human-to-robot transfer scaling laws — real human operation video is itself a pre-training asset.
Learning how scenes evolve with interaction: how objects respond to contact and how states change. Training needs multi-view spatial information and operation sequences with real causal structure.
| # | Paradigm | Key Data Need | Our Support | Recommended Datasets |
|---|---|---|---|---|
| 01 | VLA Models | Aligned vision-language-action trajectories | Full annotation: frames + action primitives + instructions | Industrial Line Ops · Home Service Interactions |
| 02 | World-Action Models | Large-scale first-person human video | 10,000h+ stock, scalable by task family | Multimodal Stock Library |
| 03 | World Models | Scene evolution & physical interaction | Stereo network + state-change annotation | Retail Shelf Ops · Robot Arm Teleop |
| 04 | Post-training / Fine-tuning | Small amounts of high-quality robot data | Pilot → batch delivery with QC reports | Custom Collection & R&D |
Head-mounted mono rigs closely match human operation recording, ready as first-person pre-training corpora.
Multi-camera networks remove blind spots, supporting spatial understanding and reconstruction tasks.
Action primitives, temporal segmentation, task instructions and state changes — annotation specs open for review.
6 own bases and a multi-form device fleet, scaling by task family with controllable delivery.
Tell us your model paradigm and training stage — we will recommend stock datasets or design a targeted collection plan.
