Tidel.ai
Capability · Robotics

Embodied data
from real hardware.

We operate the robots ourselves — teleoperation on dual-arm and humanoid platforms, tactile-glove motion capture, and egocentric vision — and deliver it synchronized, segmented, and quality-gated.

6,500+ task hours·multi-platform fleet·frame-level skill labels
How we approach embodied data

Three principles.

philosophy before catalog
01

Real hardware, real signal

Every episode is collected on physical platforms we operate — synchronized multi-camera video, depth, end-effector and joint states, and tactile signal captured together, in time. Ground truth is measured from the hardware, never estimated after the fact, so the physics in the data is the physics of the world.

02

Operators and QC in the loop

Trained teleoperators execute against task protocols, and every episode passes multi-stage human and automated review before it counts. Acceptance is a gate, not a spot check — episodes that drift from protocol, drop sync, or miss the task are rejected, and billing is on qualified data only.

03

Coverage across the pyramid

Embodied capability is built in layers — from atomic manipulation skills to long-horizon household, industrial and outdoor tasks. We design collection as a pyramid, not a pile: skill distributions are planned, tracked and reported, so diversity is a property of the dataset rather than an accident of collection.

Methodology

How an episode is made.

the same gate on every platform
Step 1 · Capture

Synchronized by construction

Multi-camera video, depth, pose and tactile streams are captured on a common clock, against a written task protocol — so alignment is a property of the rig, not a post-processing step.

Step 2 · Segment

Labels down to the frame

Episodes are segmented into skill spans at frame granularity — reach, grasp, place, recover — so the data supports skill-level training and evaluation, not just end-to-end imitation.

Step 3 · Accept

QC before it counts

Multi-stage human and automated review checks protocol adherence, sync integrity and label quality against contractual acceptance targets. What fails is re-collected, not shipped.

capture manifest · episode 3121● accepted
platform  wheeled dual-arm · teleoperated
video     4 × rgb · synchronized · common clock
depth     aligned per frame
pose      end-effector + joint states
tactile   glove array · high-rate
segments  frame-level skill spans
qc        multi-stage review → accepted
# ground truth measured from hardware — never estimated
Annotation depth

Labels you can verify.

every label traceable to a signal or a frame
Layer 1 · Structure

Episode → subtask → phase

Each recording is segmented into per-object episodes, pick / place subtasks, and motion phases — approach, grasp, lift, transport, release, retract. Phase boundaries are derived from the high-rate gripper joint signal, so every cut in the label sits on a measurable event, not an annotator's guess.

Layer 2 · Grounding

Boxes, cells and occlusion

Keyframes carry bounding boxes for every object and container in the head camera, placement is scored against the task's target cell, and per-segment occlusion is measured — so language, pixels and outcome stay attached to each other throughout the episode.

Layer 3 · Verdicts

Success is confirmed, not assumed

Each subtask ships an evidence checklist — gripper closed, object held with stable width, placed in the target cell, gripper retracted — and the verdict is recorded with the frame at which the last piece of evidence lands. Motion quality metrics (speed, path length, smoothness, pauses) ride along per phase.

label record · subtask E4 · place✓ confirmed
instruction  "Place component B in the lower-right
             cell of the sorting tray."
frames       1404 – 1621 · 7.21 s · 30 fps
phases       transport · release · retract
gripper      open @ f1585 · from 123 Hz joints
placement    lower-right cell · centred
occlusion    L 1% · R 19% max
verdict      success · confirmed @ f1619
# boundaries derived from the gripper signal — not drawn by hand
26 keyframes · head camera overlays · one real episodecomponent Acomponent Bbins
Contact sheet of 26 annotated keyframes from the head camera: bounding boxes on components A and B and on both bins, across a full sorting episode
Object boxes at every subtask boundary and every gripper event — grasped objects verified in the wrist camera, arm-occluded frames flagged. The same episode ships with synchronized video, joint states and the full label record.

A live viewer for this episode — synced video, phase timeline and per-frame label inspector — is part of sample access. Request samples.

What we collect

Collection programs.

four programs · one data pyramid
Program 01

Teleoperation

Task-directed manipulation on wheeled dual-arm and humanoid platforms — household, kitchen and industrial scenarios executed by trained operators against written protocols.

dual-armhumanoidmanipulationlong-horizon tasks
Program 02

Tactile & motion capture

Glove-based capture of human hands doing fine-grained work — contact, force and pose signal for the dexterity layer that vision alone can't teach.

tactile arrayshand posefine manipulation
Program 03

Egocentric vision

First-person video of humans performing everyday and expert tasks — the perception-and-intent layer of the pyramid, aligned with the same skill taxonomy as the robot data.

first-person videoeveryday tasksskill-aligned labels
Program 04

Field & outdoor operation

Navigation and delivery in unstructured outdoor environments — weather, terrain, pedestrians — where the long tail of the physical world actually lives.

outdoor navigationdeliveryunstructured scenes
Going deeper

From overview to catalog.

detail is disclosed progressively — by design
Layer 01 Overview

This page — platforms, methodology, and what differentiates the collection.

Layer 02 Samples

Representative episodes with full modality streams and labels, on request.

Layer 03 Pilot

A scoped collection run against your protocols and acceptance criteria.

Layer 04 Live catalog

The full browsable robotics catalog — episode inventories, scale and specs — customer access.

The full catalog — episode inventories, scale and specs — is available to customers. Request access and we'll respond within one business day.

See the data behind this page.

We'll share representative episodes, collection protocols, and scoping for custom collection programs.