Upload Modes and Storage Formats
Upload mode, storage format, and device SN are three separate decisions made at different levels:
| Dimension | Determined by | What it controls |
|---|---|---|
| Upload mode | Selected when creating the task | How data is sent back to the platform |
| Storage format | Set automatically by device model | The format in which data is ingested |
| Device SN | Selected when creating the task | Which physical device the data comes from |
Three Upload Modes
| Mode | How it works | Advantage | Trade-off | Best for |
|---|---|---|---|---|
| Full upload | Uploads every episode in full | Most complete data | Highest bandwidth use | Small batches of high-value tasks |
| Screened upload | The client filters out clearly invalid segments before uploading | Saves bandwidth and storage | Relies on collector judgment; valid data may be dropped | High-volume tasks with experienced collectors |
| Segmented upload | Splits recordings by duration and uploads while recording | Suited to long sessions; little is lost if interrupted | A segment boundary may cut through an action | All-day operation and continuous operation data |
All three modes are available for every device and are independent of the device model.
Full Upload
- Each episode is uploaded in full when it ends.
- Failed uploads are retried automatically. If the retries also fail, the episode is flagged in the task details and can be re-uploaded manually.
- No extra configuration is needed.
Screened Upload
- Before uploading, the client applies screening rules to filter out clearly invalid segments, such as all-black video or large gaps in joint data.
- Screening checks every frame, so overall upload is slower than full upload.
- Screening rules are configured and validated by R&D engineers for each customer scenario. The platform does not ship generic thresholds.
Screening is not a substitute for curation
Screening is only a coarse filter. Data uploaded with screening still needs quality labeling and human review in a curation project.
Segmented Upload
- Long recordings are split into fixed-length segments and uploaded one by one. The minimum segment length is 5 minutes.
- Uploads resume after interruptions, and duplicate or out-of-order uploads never create duplicate samples.
- When an action spans two segments, the platform stitches it back into one sample during processing.
Choosing a Mode
text
Does the task run all day?
├─ Yes → Segmented upload
└─ No → Is this a small batch of high-value data?
├─ Yes → Full upload
└─ No → Are the collectors experienced and consistent?
├─ Yes → Screened upload
└─ No → Full uploadTwo Storage Formats
| Storage format | Devices | Characteristics | Curate directly? |
|---|---|---|---|
| HDF5 | Station Lite, MOS (teleoperation) | Robot state (joints, grippers, end-effector poses) and multiple camera streams, aligned by timestamp | ✓ |
| RAW | Robot-free Ego, Robot-free UMI | Raw sensor data (images, IMU, and so on) without robot state | ✗ Pipeline required |
- The storage format is set automatically by the device model. No format option is offered when creating a task, and it can't be changed afterward.
- The format column in the asset list shows only HDF5 or RAW.
- HDF5 is also one of the dataset export formats, but the two belong to different stages: the storage format is determined at collection, while the export format is chosen when exporting a dataset.
See Data Formats for file layouts.
Bandwidth Guidelines
These are typical figures for a single device collecting continuously. Actual values depend on the number of cameras, resolution, and compression settings.
| Device | Data per 1-minute episode (approx.) | Recommended uplink |
|---|---|---|
| Station Lite | 150–250 MB | ≥ 20 Mbps |
| MOS | 200–350 MB | ≥ 30 Mbps |
| Robot-free Ego | 100–200 MB | ≥ 20 Mbps |
| Robot-free UMI | 80–150 MB | ≥ 10 Mbps |
When bandwidth is insufficient, data is cached on the device and queued for upload, so collection itself isn't affected. Task progress counts only completed uploads.

