Global Robotics Intelligence Platform
Teaching robots
the way humans
actually work.
Grip is the robot-ready data supply chain for physical AI. Capturing, enriching, and delivering high-quality egocentric human-task data at the scale foundation models need.
The gap
Robotics is missing 95% of the data it needs to scale.
Today's global data collection sits at roughly 1 million hours (about 10 human lifetimes). Frontier model teams estimate 100+ million hours are still needed to reach a true ChatGPT moment for physical AI.
At current rates, closing that data gap will take decades. Most efforts trade quality for volume, lean on teleoperation and synthetic data, or lack the production pipelines that model teams need at scale.
The platform
An end-to-end data supply chain, from human hands to robot-readiness.
Capture
Real human egocentric recording in factories, homes, and workplaces — grounded in the environments robots will actually operate in.
Quality
Production-grade standards co-designed with frontier model teams through our DM0 partnership.
Enrich
Full-stack QC, annotation, and enrichment pipelines built for physical AI, not repurposed from web data.
Deliver
Robotics-native formats — LeRobot, MCAP, ROS, and custom schemas — ready to plug into model training runs.
The product
What the delivered dataset contains
The package is designed as an inspectable robotics dataset rather than a video-only demo. Every layer is synchronized around the same episode id.
Raw video
1920 x 1080 egocentric MP4
Primary observation stream for manipulation context.
Sensor JSON
IMU / capture metadata
Synchronizes motion signals with visual frames.
Depth video
512 x 288 dense preview
Human-readable inspection layer for scene geometry.
SLAM JSONL
Per-frame camera pose
Supports trajectory review, reconstruction, and point-cloud projection.
Depth samples
Frame-level .npy depth maps
Feeds metric inspection, 3D sampling, and volume estimation.
Hand labels
3D joints and transforms
Captures interaction geometry that plain RGB misses.
Annotation
Action, subtask, skill boundaries
Turns long videos into reusable robot-learning episodes.
The demo
Takeaway box packaging and goods placement
RGB
Original raw egocentric video clip.
Depth
Per-frame metric-style depth preview.
Hand Pose
Rendered hand-pose labels over RGB.
The distribution
Multiple domains, scenes and task coverage areas.
The data sample format is consistent across domains, while the corpus emphasizes everyday unstructured spaces and repeatable fine-grained manipulation tasks.
The scale of our impact
Annotated hours of production-grade egocentric data already delivered
Hours added every month across global collection locations
Active co-design partnership shaping quality and schema standards
The advantage
Built for the teams defining physical AI.
Grip is the only player combining large scale data inventory, model-team-defined data quality standards, and a global production pipeline, at competitive unit economics for data procurement budgets.
The next era of physical AI
will be built on Grip.
If you're training embodied models, we should talk.