The biggest gap in training robots isn’t compute. It’s real-world, first-person data.
RoboX is a mobile app where contributors record egocentric video across data campaigns. Every clip gets annotated and packaged into structured datasets for robotics teams.
Join as a contributor, request API access, or launch your own custom data collection campaign at robox.to
RoboX has been accepted into NVIDIA Inception Program.
Access to @nvidia technical resources will support the expansion of our dataset infrastructure for embodied AI.
RoboX provides end-to-end data collection for robotics. Following customer capture briefs, contributors record first-person smartphone video, which our platform processes, annotates, and exports as LeRobot episodes.
Our evaluation and annotation models run on NVIDIA L4 GPUs with CUDA and cuDNN. Through Inception, we plan to extend the pipeline with Isaac GR00T, Isaac Lab, and Cosmos, so datasets can be tested in simulated policy training before they reach a customer.
RoboX has signed its first contract covering more than 1 million hours of egocentric video training data, collected for frontier robotics labs.
A year ago none of this existed. No app, no contributor network, no pipeline. What we had was a conviction that robotics was about to hit a data wall, and that the only way through it was to make data collection something ordinary people could do from anywhere.
That wall is here now. Humanoid and general purpose robots are moving from research into deployment, and every frontier lab has hit the same constraint. You can scale models and you can manufacture hardware, but demonstration data has to come from humans doing real tasks. Right now that data barely exists.
Our answer was a distributed collection network instead of a lab. It runs on two things we built from scratch. The RoboX app turns any smartphone into a capture device: contributors record everyday household and workplace tasks from a first person perspective and get paid for accepted hours.
RoboX Intelligence, our evaluation layer, reviews every recording for quality and task accuracy before it counts toward a delivery, so labs receive data they can actually train on rather than raw footage. The recordings cover cooking, cleaning, laundry, repairs and hundreds of other manipulation-heavy tasks, exactly the activities robots will be expected to perform. Today more than 1,500 contributors across Southeast Asia, LATAM and Africa are recording, with another 35,000 on the waitlist.
For context on scale: language models had the entire internet to train on. Robotics has nothing like that for human manipulation, and building it will take hundreds of millions of hours of demonstration data. We believe we are at the beginning of a multi-year growth story for the robotics data vertical.
This contract is the first proof at scale, and our collection capacity is built to grow a lot from here.
If you are training robot policies and need egocentric demonstration data, get in touch.
If you want to get paid to record, download the app and join the waitlist.
The RoboX app is now available on Google Play.
Android users can record real-world tasks, contribute egocentric data for robotics imitation learning, and earn rewards.
The app is open for everyone to install. Task participation is currently limited to users on the waitlist, and we are onboarding new contributors in batches.
Already on the waitlist? Check your inbox to see if you have been approved.
Download: play.google.com/store/apps/d…
Join the waitlist: robox.to/waitlist
RoboX LeRobot Dataset is now available on @huggingface. It loads natively in the LeRobot ecosystem and opens in the LeRobot visualizer.
This is a demonstration release: it shows that RoboX egocentric recordings convert into the standard format, and it is not ready for direct robot-policy training or deployment.
The dataset holds 28 demonstrations across 15 tasks in four domains: object transfer, organization, assembly, and kitchen work. Each demonstration is an episode with a natural language task description, egocentric RGB video, and LeRobot-compatible state and action fields of 15 dimensions covering wrist position, wrist orientation, and per-finger curl and pinch. The action field is a derived human-hand motion proxy, not a robot actuator command.
Future iterations will retarget this motion onto robot embodiments for imitation learning.
huggingface.co/datasets/Robo…
The RoboX Android Beta is now live.
RoboX is an egocentric video data collection platform for robotics imitation learning.
The app records first-person video of everyday manipulation tasks, creating the kind of demonstration data physical AI and embodied AI models use to learn how to act in the real world.
We have onboarded an initial group of beta testers. If you have already signed up, keep your notifications on so you know when your access is ready.
More updates soon as we expand beta access.
EgoTask is a data collection campaign for robotics training. First-person video of everyday manipulation tasks, recorded on an iPhone or GoPro, one to five minutes per session.
During recording, RoboX tracks both hands and labels every grasp, then turns the footage into training data for imitation learning: hand poses, grasp types, camera trajectory, and object contacts.
EgoTask dataset is now available on @huggingface.
It is a set of first-person clips of hand and object tasks, recorded through RoboX with ARKit sensors for robotics imitation learning. Tasks include making tea, loading a dishwasher, cleaning a sink, and handling objects.
Each clip provides RGB video with per-frame annotations: hand keypoints (21 joints, 2D and 3D), 6DoF camera pose, IMU, depth metadata, object tracks, and action segments. Clips include a motion-quality score and predefined train, validation, and test splits.
Research use under CC-BY-NC-4.0.
huggingface.co/datasets/Robo…
EgoTask is a new data campaign on RoboX.
A campaign for capturing complete action sequences up to 5 minutes long, recorded on iPhone. Contributors film full task executions from a first-person perspective, ideally with the phone head or chest mounted for stability.
Robotics systems learn best from complete action sequences. EgoTask captures the full arc of everyday tasks, including setup, intermediate steps, object interactions, and completion, producing data that supports imitation learning and long-horizon task understanding.
A separate campaign for external cameras is coming next.
EgoDepth is a new data campaign on RoboX.
A campaign for capturing real-world spaces in 3D using LiDAR-enabled iPhones. Contributors slowly move through indoor environments or scan objects and surfaces to build depth-aware datasets for robotics and spatial AI.
Unlike traditional navigation datasets, EgoDepth captures not only movement, but also the geometry of rooms, obstacles, furniture, and reachable space. The recordings help robots learn spatial understanding, obstacle avoidance, scene reconstruction, and depth-aware interaction with the physical world.
We’re at the final step before public release on the App Store.
For the past months, RoboX ran as a private TestFlight beta with a selected contributor group.
That phase produced over 40 hours and 15,000 episodes of egocentric data across grasping, manipulation, daily activities, navigation, and social interaction.
The public release is the start of a much larger collection phase, open to anyone with an iPhone and other supported camera devices.
EgoHub is a new campaign type on RoboX for cross-device egocentric capture.
GoPro, smart glasses, head-mounted rigs, DSLRs, and action cams are all supported, with per-device capture briefs that specify resolution, framerate, and mount position.
Every upload runs through the same annotation pipeline and feeds the same dataset structure we publish on @huggingface.
We just crossed 30 hours of real-world egocentric video.
Contributors worldwide are recording first-person footage from their phones in kitchens, offices, grocery stores, on transit. Our pipeline turns it into annotated, research-ready datasets for imitation learning and VLA models.
Live:
- 30+ hours collected across grasping, daily activities, navigation, scene understanding
- 30 to 60 fps RGB, IMU at 10 to 12 Hz, 6DoF pose via ARKit
- Depth on ~71% of clips
- On-device face blurring
- Auto-annotation: bounding boxes, hand pose, narrations, action segmentation, tracking
- Contributor reward system live
Coming:
- In-app campaigns for dataset-on-demand collection
- Expanded delivery formats for common robotics training frameworks
Scaling toward a larger, more diverse dataset across geographies, environments, and task types that matter for embodied AI training.
Our latest egocentric video dataset is now available on @huggingface.
7,342 annotated first-person clips for robotics imitation learning, organized into four campaigns: grasping, daily activities, scene captures, and navigation.
Every clip includes hand keypoints, object tracks, action segments, 6DoF camera pose, and 200Hz IMU.
Free for research under CC-BY-NC-4.0.
huggingface.co/datasets/Robo…
EgoGrasp sample dataset is now live on @kaggle.
Everyday object grasping captured from a first-person view, with layered annotations for hand pose, object tracking, and action segmentation. Ready to explore and build on directly in a notebook.
Interested in the full dataset? Request a pilot at robox.to.
🔗 kaggle.com/datasets/roboxtec…
We have selected the first group of contributors to access the beta version of the RoboX app.
If you signed up for early access, check your email for an invitation. The message includes setup instructions and details on how to complete your first recording session.
We are opening more spots in the coming weeks.
If you have not signed up yet, you can join the waitlist at robox.to/waitlist.
EgoGrasp sample dataset is now available on @huggingface.
First-person video clips of everyday object grasping with layered annotations covering hand pose, object tracking, and action segmentation.
Interested in the full dataset?
Request a pilot at robox.to.
🔗 huggingface.co/datasets/Robo…
We collected 1,800 egocentric grasp episodes during private beta through the RoboX app.
620+ unique objects recorded from a first-person perspective across real environments.
Annotated and ready for humanoid training pipelines.
Request a pilot dataset or scope a custom collection campaign at robox.to
A manipulation policy doesn't train on video. It trains on video paired with 21-point hand pose sequences, per-frame object bounding boxes with in-hand status flags, action phase segmentation, grip classification, and natural language task descriptions, all synchronized to the same frame index.
This article walks through how RoboX annotates first-person video at scale and how frame-level alignment across all layers produces training samples ready for direct ingestion into policy learning workflows.
robox.to/blog/inside-the-rob…
Robots have learned to walk, carry boxes, and navigate warehouses.
They still can’t do what humans do every morning without thinking.
The gap is egocentric data, first-person recordings of how humans see, grasp, and navigate the world.
The most valuable and least collected dataset in robotics.
Every dataset we build at RoboX starts from a single perspective, human.
















