Visual intelligence lab

The Video
Datalake for physical AI, robotics, world models, VLAs and humanoids.

Egocentric video, labeled by our own models.

Egocentric data: Our own rigs at real work sites. 1K+ tasks across 20+ industries.
Dense annotation: Your teleop, UMI, egocentric or sim video, per your taxonomy.
EMG force data Coming soon: Wrist EMG decodes contact and force, synced to the video and hand pose.

Explore Sample

Left hand: The hand holds the dosing cup under the grinder spout, steadying it against the chute while the grounds fall in.Right hand: The hand reaches to the front of the grinder and rests against it beside the other hand, fingers loosely curled.

Dosing and weighing ground coffeeRight hand: Reach outLeft hand: Manipulate in hand

Actions
Captions
Hand pose
Raw clip

investors and partners

Video data infrastructure for the frontier.

We're building a curated Video Datalake with millions of hours of high-quality egocentric video data and annotation services for our customers pushing the frontier of physical AI.

Real work, everywhere.

Our own rigs record at real work sites: a head-mounted stereo pair and a wrist view, synced.

Filtered. Indexed. Growing.

Every clip passes QC and is indexed per hand, action and task before it joins the lake.

Last monthThis monthNext month

Ready to train robots.

Clips arrive in the shape robot learning needs.

Robust infrastructure and network.

We run the end-to-end system that enables efficient training cycles.

Capture

Our sites, our rigs.

Wearable rigs record video where the work happens, mounted at the head, wrist, chest, cap or neck. Also captured: raw IMU and camera calibration.

4-camera headset

Stereo headband

Wrist camera

Index

Every action, indexed per hand.

Each hand's actions become time-stamped segments, so the lake can be searched by action.

Explore

Left and right hand · action segmentsClip S8 · 06:54 → 07:19

L · GraspR · Grasp

Retrieve

Curate what's usable.

We query the lake for the video you need, then annotate it against your taxonomy.

Read the API docs

# retrieve.py · illustrative
moments = lake.retrieve(
    venue="construction site",
    scenes=["rebar area",
            "formwork area"],
    skill_level="hard",
    where="both hands on the tool",
    exclude=["hand idle", "walking"],
)
moments.annotate(
    taxonomy="your_taxonomy.yaml")
# → rebar and formwork moments, labelled

Annotate

Dense captions, to the frame.

Each hand gets a caption for every segment, written against your taxonomy frame by frame.

Explore

First-person view at an assembly station: gloved hands drive a screw into a black part with an orange electric screwdriver, bins of parts and screws beside them

07:00.043RightThe hand holds an item stationary in a pinch grip, then shifts slightly inward while adjusting its fingers around it.

07:01.376LeftThe hand holds the item stationary in front of the body, slightly shifting its thumb and fingers to maintain its grip.

Deliver

Annotations as JSON, per segment.

Every segment ships with its hand, label, start and end time and caption, ready for training.

See the delivery format

# s8-segment.json · excerpt
{
  "clip": "S8",
  "track": "right_hand",
  "label": "Grasp",
  "primitive": "grasp",
  "start": 427.079,
  "end": 430.876,
  "text": "The hand holds the component…"
}