Robot training data, recorded in the real world and labelled to your rules.

Egocentric video and other real-world robot data from our partners, checked recording by recording and delivered ready to train on. Use collection, annotation or both.

Request a pilot See collection

action: pour

Partners

TanahLabs

VR Oxygen

Records: 1.5M+
Hours recorded: 11,900+
Total size: 76 TB+
Partners: 4

Problem

Robot models are short of real-world data, and the data they have waits on labels.

Language models learned from the internet. Robots have no internet of physical work to learn from. Most open robot datasets come from a handful of labs, recorded on the same benches, under licences that rarely cover commercial use.

Teams that need real homes, warehouses and kitchens end up collecting the data themselves, then labelling it frame by frame.

One frame: a robot hand tilting a bottle over a cup, with boxes around the hand, bottle and cup. — At 30 frames a second, with three tracked objects in every frame.

What we do

Two products, used alone or together.

Collection

Demonstrations recorded to your task list.

Partners record in homes, workplaces and sites you choose, to the task list you set.

Our partners cover egocentric video, real-home demonstrations and fixed-camera data.

See collection

Head-camera frame of a robot hand tilting a bottle over a cup, with labelled boxes around the hand, bottle and cup.

Annotation

Your recordings, labelled in your schema.

Send video from any robot or rig; Talo returns boxes, tracks, action segments and task descriptions.

Each batch ships with the labelling rules and a check log.

See annotation

Our data

We turn everyday work into robot training data.

Task video

Head-mounted and overhead video of everyday work in homes, kitchens, markets and workshops, with step labels.

Robot teleoperation

An operator drives your robot through your tasks, with joint states and commands recorded in sync.

Fixed-camera sites

Cameras at working sites, such as waste disposal points, with each item's type and weight recorded.

Handheld grippers

Wrist-camera video with the gripper's position and opening width, recorded without a robot.

How it works

You check the data fits before anyone records at volume.

Write the task list

Each task becomes a short script: start state, end state and what may vary. Already have footage? Send it and skip to step 4.

task:     pour water
start:    bottle upright on the table
end:      bottle back on the table
may vary: cup position, fill level, room

Test a pilot week

A few tasks in a few locations. You load them into your training code to check the data fits before any larger order.

scope:    a few tasks, a few locations
time:     one week
your job: load it, train on it, decide

Partners record

In homes or sites matched to your list, on the capture rig that suits the task.

capture:  head camera | phone | fixed camera
sites:    matched to your task list
consent:  on record for every contributor

Check and label

Every recording passes an intake check, then is labelled to the agreed rules. This is our public sample clip.

tracks:   right hand, bottle, cup
segments: grasp, tilt, pour, return
check:    28 frames, 9 corrections logged

Receive weekly batches

Batches in your format, with the rules and a check log, so you can see how every label was made.

format:   your schema; Parquet and MP4 by default
rules:    included with each batch
check:    log included with each batch
access:   private, access-logged repository

Quality

You can see how every recording was vetted and every label was made.

Vendors hand back files. Talo hands back files and the record behind them, so you know which data to trust and can cite how it was made.

Check a labelled sample

Consent on record
Each partner runs its own consent process, listed on its partner page.
Intake check on every recording
Recordings that miss the task script or the capture spec do not reach labelling.
Rules agreed with you
Action names, where each segment starts and ends, and what counts as a failure.
Check log with each batch
Frames checked and corrections made, delivered with the labels.

Case studies