Dexterous manipulation
Hand poses and object poses become grasp labels for a policy.
From our notes
All notes →Research, releases and milestones
Integrations
All models →Every model, one action layer
Run leading video, perception, 3D and policy models on your captures — side by side, on the same episodes.
Integrations
Bring your data. Keep your stack.
Import from wherever your footage lives, process it with any model — yourself in Studio, or with the Actions agent over MCP — and send the datasets straight back.
01 · Import
Hugging Facehf://datasets/…
Amazon S3s3://robot-demos/
Google Cloud Storagegs://fleet-logs/
Google DriveShared drives
Local uploaddrag & drop .mp4
02 · Process with any model
cooking.mp4sorting_packages.mp4tea_making.mp4
Segmentationdone
Hand posedone
Depth map35%
Captions0%
Runs on+ your own
03 · Export
Episodes, frame-aligned
RGB videomasksdepthhand poseactionscaptions
AsJSONepisode recordsParquetcolumnar tablesMCAPROS / Foxglove logsHDF5numeric arraysWebDatasetsharded .tar
Label anything
Every reality, fully labeled
Depth Map, Segmentation, Hand Pose, and Analyze on every frame, with masks and the rest of the pose set on the way.
A pipeline for every use case
Chain models over the same capture. Outputs line up frame by frame, so each step feeds the next.
Humanoids
One demonstration, transferred across scenes and replayed in sim.
Egocentric video
Head-mounted footage segmented, tracked and described step by step.
Warehouse picking
Find any SKU by name, measure it, and score picks before deploying.
Home robots
Rebuild the room, generate new versions of it, fine-tune an open VLA.
Policy evaluation
Roll a policy forward in a world model and grade it with a VLM.
One capture · every world
Your world, multiplied.
Bring your captures and your models. We turn them into every version of the world your robot will meet — labeled, simulated and ready to train on.