Robot Data

Training data for robot foundation models

Synchronized capture from the worker’s point of view and from the room, annotated for what the model needs to learn. Every clip was recorded in a working environment. Customer names are withheld by agreement.

( 01 )

VLM datasets

Robot training data

Multi-view video paired with questions and answers about what is happening in the scene. Built to test spatial reasoning, physical reasoning, and embodied action.

Capture
Ego + exo, synchronized
Annotation
Capability-probed Q&A pairs
Format
Video with paired Q&A. Schema agreed per program.

A sample of the VLM library. The PRISM-100K open subset is on Hugging Face. Full set under NDA.

( 03 )

Loco-manipulation datasets

Robot training data

Work that involves moving through the space, not just working at a station. Approach, reach, whole-body positioning, and two-handed work at close range, captured from the worker’s view and from the room at the same time.

Capture
Ego + exo + wrist, synchronized
Annotation
Hand pose, contact, action boundaries, task breakdown
Format
LeRobot (RLDS), Open X-Embodiment
Stocking and replenishment
Grocery · ego + exo + wrist
Pick and pack
Retail store · ego + exo + wrist
Food prep, knife work
Commercial kitchen · ego + exo + wrist

A sample of the loco-manipulation library. Full dataset available under NDA.

How a capture program works

Capture
Exo
Alia records the room. A single-shot 360° stereo panorama with depth, from one camera position.
Ego
A head-mounted camera on the worker, synchronized to Alia frame for frame.
Wrist
Added for close manipulation. Your own cameras can be included.
Annotate
Method
AI does the first pass. A person reviews every batch. Segmentation, tracking, action boundaries, and task breakdown are set per program.
Deliver
Format
In the format your training stack uses: LeRobot (RLDS), Open X-Embodiment, or a schema we agree before capture starts.

Quality

Four gates

Every capture passes four gates before we deliver it.

G1  Calibration
Checked every session.
G2  Alignment
All camera streams confirmed in sync before annotation starts.
G3  Annotation review
A person reviews every batch.
G4  Rejection
Batches below threshold are captured again.

Full numeric specification available under NDA.

Where we capture

We have agreements and compliance approvals in place across grocery, pharmacy, hardware, fashion retail, restaurant kitchens, manufacturing, warehousing, and clinical and home care settings.

Tell us what your model needs to learn

Capture programs, research collaboration, and dataset partnerships.

Talk to us