Pick up the cup
271f·18s·Stereo depth cam
Reference episode · MANO 21-kpt hand + verb-noun label
Industrial data · real production environments · not scraped, not simulated.
The Robotics Data Foundry for Physical AI
Real capture packs, worn by operators on the factory floor — egocentric, action-paired, deeply annotated. Sourced through an industrial partner network across Asia, QC'd, and delivered RLDS-ready.
Egocentric captureAI-native QCRLDS-ready
Egocentric captureAI-native QCRLDS-readyIndustrial network · real production · LeRobot / RLDS · ≤48h delivery
THE PROBLEM
Physical AI is blocked by real-world data — not compute.
Robot foundation models don't lack compute. They lack synchronized, action-paired data captured in the messy real world — and every lab is hitting the same wall.
FIG.00 — THE DATA GAP
● Model & compute scale● Real-world data
Compute & model scale race ahead; real-world, action-paired data lags. That gap — not compute — is the bottleneck.
10×
More demos per hour from egocentric human video than teleoperation
EgoMimic · Georgia Tech
log-linear
20k+ hrs of egocentric video → predictable gains in robot dexterity
EgoScale · NVIDIA
<62 hrs
Robot data + 1M hrs of web video → zero-shot manipulation
V-JEPA 2 · Meta
~10,000 hrs
π0: data across 7 platforms trains one cross-embodiment VLA
Physical Intelligence
The scarce resource is real, diverse, action-paired demonstrations — QC'd to lab standard.
THE FOUNDRY
So we built a foundry for it
Tbrain is a robotics-data foundry: we capture real, action-paired demonstrations on our own hardware and forge them into lab-grade, training-ready datasets.
SOURCE · VIETNAM PARTNER NETWORK
Garment factories · 1,000+ operators
Auto garages & repair workshops
Electronics assembly lines
Commercial kitchens & food service
Warehouses · markets · retail
Tbrain · the data foundry
CAPTURE · SYNC · QC · ANNOTATE → RLDS
Collect01
Sync02
Annotate03
QC04
Deliver05
POWERS · ROBOT FOUNDATION MODELS
VLA / foundation models
World models & neural sim
Humanoid robots
Manipulation & RL policies
FIG.01 — THE DELIVERY SPEC
Built to the spec world-model teams ask for
Every episode ships to one spec — the exact signals world-model teams request, hardware-clock synced, QC'd, and delivered RLDS / LeRobot-ready.
Egocentric POV
First-person — the exact view a robot acts from.
Action-paired
Every frame aligned to the action / teleop signal.
Hardware-clock sync
Sub-frame alignment across all sensors.
Multi-modal
RGB · Depth · IMU · Audio in one record.
RLDS / LeRobot-ready
Drops straight into your training stack.
QC'd · ≤48h
AI-native quality control, fast turnaround.
RECORDING · REAL PRODUCTION
FIG.02 · CAPTURE WALL · 6 CONCURRENT SESSIONS
Real production.
Zero staging.
Every clip on this wall is a real capture from a real factory floor — auto-labeled, QC'd against 15 machine-checkable rules, and diffable against its Rerun scene.
0
tasks
0
frames · sample
0+
caps · queue
15/15
hard rules PASS
92%
first-pass ship
SEW · TEXTILE
LIVE
ARRANGE · TEXTILE
LIVE
PACKAGE · TEXTILE
LIVE
PICK · TABLETOP
LIVE
EXO · MOCAP
LIVE
status● 17 tasks captured
6 sample clips · 17 tasks total·4 ego + 1 tabletop + 1 exo·consent signed · faces off-frame
FIG.04 — WHAT SHIPS NOW → NEXT
20 skills shipped · 500-pack fleet next
One rule: verifiable numbers on the left, aspirational capacity on the right — labeled, not blended. We don't ship stats we can't defend to a research engineer.
SRC · model registry · LeRobot export · capture ledger
Industrial partner network across Asia — factories, workshops, assembly lines
Industrial
Real factory floors · not staged
Multi-site
Textile · kitchen · electronics roadmap
LeRobot / RLDS
Ready for the training loop
≤ 48h
Raw → QC'd → delivered
0
Models · auto-label pipeline
Hand · body · masks · depth · verb-noun
0
Hard rules · per capture
Machine-checkable QC gate
0%
Ship rate · after QC
Auto-accept + Label Studio + reviewer
0h
Delivery latency
Raw → LeRobot v2 · zero-trust
FIG.04 — FOUNDRY PIPELINE
One pipeline · five phases · every stage diffable
From factory floor to LeRobot v2 in ≤48h. Every phase leaves a machine-readable trace so any downstream claim can be verified.
IRON · TEXTILE
REAL
iron_product · op mobile · 450 f
SEW · TEXTILE
REAL
sew_hem · op mobile · 450 f
ARRANGE · TEXTILE
REAL
arrange_fabric · op mobile · 450 f
PICK · TABLETOP
REAL
pick_up_the_cup · op unknown · 273 f · 15/15 PASS
Real captures · zero staging · every clip auto-labeled + QC'd on this landing
● 4 clips live-looping·ego + tabletop·faces off-frame · consent signed
FIG.10 — CAPTURE LEDGER
Real captures · pick one, open its Rerun scene
Fourteen episodes shipped this quarter across kitchen + textile lines; six on this page. Every card opens its own annotated burn, its own manifest, its own Rerun scene. Click any tile.
KITCHENKitchen ledger· 8 entries
Delivered · QC-verifiedTEXTILETextile factory· 8 entries
8 raw egocentric captures · annotation downstreamFIG.05 — AUTO-LABEL · 4 SIGNATURE OUTPUTS
Per-stage deep dive
Every raw capture runs through the auto-label pipeline in parallel. Four outputs are the ones a robotics team touches first — the rest live in the provenance manifest that ships alongside the video.
FIG.05C — KEYPOINTS · HAND
Hand keypoints
Per-frame 21-keypoint MANO mesh + SLAM camera trajectory for each hand independently. Interpolated frames flagged; low-coverage caps escalated to Label Studio.
output · hands.left/right · kpts_2d · kpts_3d_world · kpts_3d_cam · source
RAW · rgb.mp4
OVERLAY · MANO 21-kpt + per-hand SLAMhands.left/right
Description, metadata, object masks, depth, Rerun proof — the full 8-model pipeline lives on the deep dive page.
See the full pipelineFIG.06 — QC · HARD RULES + AI + HUMAN
Zero-trust QC · 15 hard rules → AI filter → 3 human layers
Every capture crosses a 15-check gate before a human ever sees it. Only PARTIAL/FAIL results route into Label Studio, where three human layers ship the last 8%. Every fix keeps the provenance trail intact.
LAYER 1 · HARD RULES · 15 CHECKS · LIVElive · pick_up_the_cup 20260617T01
categories · gate composition
calibration0/2
temporal0/2
detection0/4
spatial0/3
semantic0/3
provenance0/1
● 15 · pass· 0 partial · 0 fail|reject rate global 22%|provenance · git 1b0cce1
LAYERS 2–4 · HUMAN REVIEW
- Layer 1 · Label StudioAuto-label outputs load into Label Studio as pre-populated tasks. Annotators correct kpt drift, adjust masks, override verb-noun. Every correction is a labeled diff.
- Layer 2 · Reviewer sign-offA second annotator reviews the correction. Accept · reject · flag. Rejections send the task back with reason codes.
- Layer 3 · Escalation dashboardSystemic failures (segmenter locked wrong object, SLAM divergence) escalate to engineering. Root-cause reports feed back into the auto-label training loop.
FIG.07 — SHIP-RATE DELTA
0 → 92% ship-ready after QC
Raw auto-label
0%
After hard-rules gate
0%
After AI-filter refine
0%
After Label Studio fix
0%
After reviewer sign-off
0%
Full 15-check taxonomy, sample fail images per check, escalation flow, and the models provenance trail live on the QC playbook.
See full QC playbookFIG.11 — RERUN EPISODE VIEWER
Every episode is a multi-track Rerun scene
Open any capture as a scrubbable multi-track scene. Full viewer lives on the auto-label deep dive.
9 tracks · scrubbable · v0.25
FIG.08 — BUYER LENS · SHIP · BUILD · CHECK
One page, three answers
What we ship, what you're building, what to check before you buy. Same buyer question, three lenses. Click any tile to expand its detail.
Module A · NOW
Egocentric video
VLA · world-model
First-person capture packs worn by operators on real factory floors. Every episode ships with per-frame hand kpts + object masks + verb-noun action segments + camera SLAM trajectory, all baked into one LeRobot v2 parquet.
FPS
15
Frames / ep
180–540
Latency
≤ 48h
Export
LeRobot v2
LeRobot v2 fields · per episode
Sample captures
pick_up_the_cup · 20260617T01iron_product · 20260626T01sew_hem · 20260626T02arrange_fabric · 20260626T01
You receive
- ✓ LeRobot v2 parquet + mp4
- ✓ Rerun .rrd (9 tracks)
- ✓ Annotated burn (palette-coded)
- ✓ Per-field provenance manifest
FIG.07 — THE RIG + THE APP
The rig and the app
Two views of one collection machine — the wearable pack and the operator's task console. Purpose-built so 500 operators capture the same schema, same QC gate, same bucket.
FIG.15 — PUBLIC REFERENCE WALL
We build on top of the open frontier
Public egocentric datasets set the ceiling for what a research team already expects. We benchmark to them, extend them into East-Asian environments simulation never sees, and deliver in the same schemas. Everything below is credited public work — not ours.
PUBLIC
Ego4D
3,025 h
Multi-domain
PUBLIC
EPIC-Kitchens
100 h
Kitchen
PUBLIC
EGTEA Gaze+
28 h
Kitchen
PUBLIC
EgoCom
38 h
Human conversation
PUBLIC
TREK-150
1.5 h
First-person tracking
Public references only · we do not resell public data · all links go to the original project pages.
FIG.16 — WHY TBRAIN
Why buyers pick our foundry
We win on annotation depth, accountability, and a field-scale operation no US-centric vendor has — not a race to the bottom on price.
Annotation depth
Hand pose · sub-action · failure/recovery · human audit
Accountability
Recruit · train · QC · QA report on every delivery
Environment diversity
Residential · commercial · industrial · retail · field
Interoperability
LeRobot / RLDS — plugs into modern training pipelines
Field scale
Vietnam university partner network · ramps in parallel
0
Parallel capture packs at scale
0+
Episodes / month at scale
0+
Environment categories
0
Synchronized streams / episode
FIG.17 — HOW TO ENGAGE
Enter anywhere
There's no big upfront commitment. Most buyers begin with inventory or a low-risk pilot, then scale into production and a retainer.
ENTRY
License Inventory
Access pre-collected egocentric data quickly.
accessfast
VALIDATE
Pilot Collection
Validate task protocol, hardware, workflow & output format.
scopeshort pilot
HARVEST
Production Program
Collect large-scale data across targeted environments.
runtimescaled program
REFINE
Retainer
Monthly data flywheel for continuous model refinement.
cadencemonthly
DATA GOVERNANCE · PROCUREMENT-READY
Private programs
by buyer spec
QA report
every delivery
Consent & usage rights
per project scope & SOW
DPA / security
on pilot & procurement
Enter at any point — license-ready inventory today, custom collection to your spec, or a continuous program.
Talk to usBEYOND ROBOTICS
Tbrain also runs coding, evaluation, and RLHF / SFT data programs.
Forge your next dataset with us
Tell us the task, the embodiment, and the format. We'll scope a sample batch — captured, QC'd, and delivered RLDS-ready.