Skip to main content
Langhui

Home / Embodied AI data services

PHYSICAL AI · DATA & CAPTURE

From real-world tasks
to testable training data

Start with what your robot needs to learn. Align data sourcing, custom capture and equipment through task definition, a pilot, annotation and versioned delivery.

Review samples and licensing before confirming volume, price and schedule.

E2 public recording · post-processing demoThis is a 4 Hz preview. Skeletons, masking and the inset are post-processing, not raw device output or training source. Source & field notes →

WAYS TO WORK

Start with your buying task

02 / COMMISSION

Commission capture

Define tasks, target embodiment, counting units and acceptance. Validate feasibility with a pilot.

Build a project brief →
03 / COLLABORATE

Equipment & capture partners

Explore TOBI equipment for your own capture, or use the partner channel for sites, expertise and capture capacity.

Compare capture equipment →

Have data, sites or specialist services to offer? Explore expert and supplier collaboration →

CHOOSE THE DATA

Choose by observations and actions

The three paths need different hardware, fields and validation.

EGO

Egocentric video

Study human workflows, scene semantics and interactions. Define observations separately from robot-action labels.

Delivery to agree
Raw video, timestamps, available IMU and calibration; derived labels listed separately.
Review focus
Mono/stereo, field of view and distortion, timing, dropped frames and contributor/site coverage.
UMI

Handheld gripper demonstrations

Record manipulation with a sensorized handheld gripper. UMI projects require feasibility review and trajectory validation for the target robot.

Delivery to agree
Wrist video, valid trajectories, gripper width and calibration; fields confirmed during a pilot.
Review focus
Single/bimanual setup, tracking failures, gripper calibration and target-robot kinematics.
TELEOP

Robot teleoperation

Record operator commands and robot states on specified hardware. Recheck interfaces and action space when the robot or control setup changes.

Delivery to agree
Camera observations, joint states, action commands, interventions and outcomes as scoped.
Review focus
Robot model, control rate, common timebase, success/failure and recovery episodes.

UMI is a public research method; project evaluation does not establish off-the-shelf availability. Method reference: Universal Manipulation Interface.

EVIDENCE BEFORE SCALE

Make the pilot testable

A specification states requirements. Samples and checks establish what was delivered.

  1. Task, environment & countingSpecify unique hours, episodes or clips. Do not multiply concurrent camera time; report raw and usable duration separately.
  2. Observations, actions & timingList fields, clock sources, timestamp definitions, calibration and the method for measuring synchronization error.
  3. Quality, failures & splitsAgree complete/invalid episodes, failure/recovery records and participant- or scene-based train/test separation.
  4. Source, licensing & permitted useConfirm participant/site permissions, privacy handling, training use, exclusivity and redistribution limits.
  5. Format, versions & acceptanceTest a sample loader; agree format version, coordinates, units, missing values, validation records and rework conditions.

ILLUSTRATIVE DELIVERY STRUCTURE

batch_manifest.json
episodes/episode_id/
  observations/
  timestamps/
  calibration/
  labels/
quality/checks.json
metadata/data_card.json
README + version notes

This is a discussion structure, not a delivered dataset. Agree fields, formats and acceptance thresholds for the project.

Review sequence: requirements and permissions → pilot validation → agreed specification and quote → batch delivery and acceptance.

BUILD YOUR BRIEF

Turn your task into a capture brief

Prepare requirements here, then add contact details on the next page to submit to our team.

Describe the business scope; do not enter patient, customer or contributor identities.

BEFORE YOU BUY

Resolve the key buying questions

Do catalog quantities establish stock?

The catalog helps match requirements. Confirm usable quantities, duration, field completeness and permissions for a specific batch.

Can a public preview be used as training data?

Public previews illustrate capture and processing. Raw files, annotations, licensing and model suitability require a separate review.

Why is there no universal hourly price?

Cost depends on tasks, contributors, sites, sensors, embodiment, usable-episode requirements, annotation, rights and schedule. Agree counting units and a pilot before pricing.

Can we use our own equipment or robot?

Submit your configuration and tasks for review. Agree interfaces, synchronization, permissions, safe operation and acceptance before defining scope.