Data Collection Overview
Data collection is where the NexCore workflow begins. The platform supports two collection methods, plus a third data source that requires no dedicated collection sessions:
| Method | Devices | What is captured | Storage format | Learn more |
|---|---|---|---|---|
| Teleoperation data collection | Station Lite (dual-arm workstation), MOS (wheeled dual-arm robot) | An operator teleoperates the robot, recording robot state, actions, and multiple camera streams | HDF5 | Teleoperation Data Collection |
| Egocentric data collection | Robot-free Ego, Robot-free UMI | A person performs the demonstration directly, recording first-person stereo video, IMU, and other raw sensor data | RAW | Egocentric Data Collection |
| Continuous operation data | Robots running in production | The full working day (teleoperation, autonomous, takeover, idle, and failure) flows back to the platform | 5-minute archives | — |
Choosing a Method
| Consideration | Teleoperation | Egocentric |
|---|---|---|
| Consistency with the target robot | High: actions are produced directly on the target embodiment | Requires pipeline processing and retargeting |
| Collection throughput | Limited by operator skill and the number of robots | Higher: people demonstrate directly |
| Scene coverage | The robot must be brought to the scene | Lightweight devices reach more real-world environments |
| Downstream processing | Can be curated directly | Must go through a data pipeline first |
Both kinds of data can be combined in one project, for example egocentric data to broaden scene coverage and teleoperation data to keep the embodiment consistent.
Collection Workflow
text
Register device → Create collection task → Collector claims task and records → Upload → Integrity check → Data assets| Step | Description | Reference |
|---|---|---|
| Register device | Add the device model and SN to the device inventory | Device Management |
| Create collection task | Select model, select SN, select upload mode, enter task details | Collection Tasks |
| Collect and upload | The collector runs the task; data uploads according to the upload mode | Data Upload |
| Validate and ingest | Data is assigned a data ID once channels, timestamps, and package counts are consistent | Data Assets |
Collection Best Practices
- Start small: Record 10 to 20 episodes, curate them, and run a short training job to validate the pipeline before collecting at scale.
- Keep demonstrations consistent: Within a task, keep the starting pose, object placement range, and pace as consistent as possible.
- Cover variation deliberately: Within that consistency, intentionally vary object positions, lighting, and backgrounds.
- Keep failures: Failed demonstrations don't need to be deleted. Flag them during curation, because failure data is valuable too.
- Document the setup: Record the scene layout, object specifications, and similar details in the task notes for later traceability.

