How RoboX Works

From Collection to Model Training

RoboX operates a three-stage process that moves data from raw smartphone recordings to structured, annotated datasets ready for robotics training.

Stage 1: Collection

How Contributors Collect Data

Contributors download the RoboX app, browse active campaigns, and start recording. Each campaign defines the task, sensors, and clip format.

Campaign-Specific Briefs

EgoGrasp asks for close-range hand-object interactions. Capture how you pick up, hold, manipulate, and place everyday items. The focus is on hand visibility, object contact, and natural grasping motion.

EgoNav asks for walking through indoor spaces like offices, malls, apartments, or airports. Capture your perspective as you move, navigate obstacles, and interact with the environment. The focus is on spatial understanding and obstacle awareness.

Contributors pick the campaigns that fit their routine. There's no minimum duration requirement, even brief, high-quality clips are valuable.

Lightweight Capture

Recording uses the phone's native camera, IMU, and where supported, LiDAR or audio. Short, structured clips are captured during normal activity. The app handles scheduling and sensor coordination so you don't have to think about it.

Stage 2: Processing and Anonymization

On-Device Anonymization

Faces, readable text, license plates, and device identifiers are detected and removed locally before anything uploads. Only clean, anonymized sensor data leaves the phone.

This happens on your device, before your data ever touches a server. You retain full control over what's anonymized and what's sent.

Quality Validation

Every clip goes through automated checks:

• Lighting: Is the clip well-lit and usable?

• Stability: Is the camera steady enough to extract meaningful signals?

• Completeness: Does the clip cover the full task or motion?

• Relevance: Does the content match the campaign brief?

Only clips that pass are accepted. Contributors get clear feedback on rejections through the app, so you know what to improve next time.

Stage 3: Annotation and Distribution

Layered Annotation

Validated clips are annotated with multiple data layers:

• Temporal segmentation: Key phases of the task (approach, contact, manipulation, release)

• Object bounding: What objects are present and their spatial bounds

• Hand pose: Joint positions and hand configuration throughout the clip

• Gaze direction: Estimated eye fixation and attention patterns

• Spatial layout: Scene structure, surfaces, and spatial relationships

• Interaction labels: Grasp type, object category, action outcome, and task semantics

New annotation layers can be applied to existing recordings as schemas evolve. Your contribution remains useful as new research questions emerge.

Dataset Access

Annotated datasets are organized by campaign and accessible via API or direct download. Teams can:

• Browse existing datasets: Search by task, environment, sensor type, or region

• Request access to specific campaigns: Get notified when new data is added

• Commission custom collections: Work with RoboX to design and execute targeted data missions with defined geography, environment, task, and sensor parameters

For Data Companies and Integrators

RoboX datasets are also available for licensing and redistribution through data marketplace partners. Companies that aggregate, resell, or integrate training data into their own pipelines can access RoboX egocentric datasets in bulk, with:

• Flexible licensing terms: Commercial, academic, and research licenses available

• Structured metadata: Ready for downstream integration and pipeline automation

• Version control: Track dataset updates, changes, and new annotation layers

• Compliance support: All datasets include privacy and provenance documentation

Contributor Experience: Step by Step

01 Select a Campaign

Open the app and browse active data missions. Each campaign has a clear brief covering what to record, how long, and which sensors are needed.

02 Record

Go about your routine while the phone captures short, structured clips during the campaign window. No special equipment required.

03 Auto-Anonymize

Before anything uploads, the app strips sensitive identifiers directly on the device, including faces, text on screens, license plates, and voice data.

04 Data Reaches R&D

Validated, anonymized, annotated clips are organized into campaign-specific datasets and made accessible to robotics teams through API or direct download.

What Happens to Your Data

Your contribution to RoboX:

• Is anonymized on your device

  Before upload, all identifying information is removed

• Is validated automatically Quality checks ensure it meets the campaign standard

• Is annotated with structured labels Making it useful for robotics training

• Is organized into public or commissioned datasets Ready for research and development

• Stays under your control You can request deletion anytime, and you can see where it's used

• Is tracked on-chain Data provenance is cryptographically verified