Embodied AI · Data Infrastructure · CROSS-BORDER · COMPLIANCE-ROUTED · ASSET-LIGHT

The data layer for the humanoid era.

Real-world capture, teleoperation, simulation, RLHF — brokered across vetted Chinese and APAC capacity. Single MSA. Compliance-routed. Money-back QA guarantee.

Humanoid teams ship slower because training data procurement still takes a quarter. Five distinct data types. Five fragmented vendor stacks. PIPL gates the largest supply pool on the planet.
01 · Fragmentation
5 vendor stacks, 5 contracts
Real-world / synthetic / RLHF / teleop / evals — each from different vendors, different geographies, different compliance regimes.
02 · Procurement
4-6 months to first delivery
NDA, vendor due-diligence, cross-border legal review, QA pilots — burning a quarter before usable data lands.
03 · Compliance
PIPL gates the largest pool
China supply is ~40% of global capacity. PIPL + data-export + US BIS gate it for foreign labs. Most leave it on the table.
04 · Pricing
$340/hr → $118/hr in 18 months
Teleop data benchmark collapsed 65% (Q1'24 → Q1'26). Labs still paying legacy rates lose 2-3× margin.
What you get

A three-piece bundle
that lands in your training stack.

Every pilot produces the same three artifacts. Each is a standalone deliverable; together they let your team simulate, train, and ship a new policy on day one of the bundle landing.

Isometric workcell · table-mounted robot arm with tripod camera observing a target object
Piece 1 · Sim-ready scene
A digital twin of your workcell.

Run simulation, evaluation, and policy testing before touching a real robot. Geometry, mass, and friction tuned to your actual environment so the sim and the real workcell agree.

Formats: URDF + MJCF + USDA · drop-in for Isaac Sim, Genesis, SAPIEN, MuJoCo

Sequence of six robot arm poses across time with a velocity profile sparkline below
Piece 2 · Trajectory dataset
Demonstrations multiplied 10× across grippers.

Real on-site demonstrations grown via jitter, time-stretch, pose-offset, and Blender re-rendering. Cross-gripper retargeted across Franka Panda, SO-100, Allegro Hand and four more so the dataset isn't locked to one embodiment.

Format: LeRobot v3 + Open X-Embodiment compatible · HuggingFace dataset card included

Flow diagram · stacked base model layers, plus a small adapter delta box, equals a task-specific deployable skill
Piece 3 · LoRA adapter
A drop-in task skill for your policy stack.

Load into your robotics stack and start evaluation immediately. Fine-tuned on your choice of backbone — OpenVLA-7B, π0, RT-X, Octo, RDT, or GR00T N1.7 — trained onshore and delivered onshore-only.

Format: HuggingFace PEFT · backbone choice at pilot kickoff

How it's made

Six steps. Same shape, every pilot.

Every Teehoo pilot runs the same six-stage pipeline. Each stage emits an audit event into the bundle's hash chain. You can verify the chain after delivery with one CLI command.

Six-stage Teehoo pipeline from left to right: Collect · Reconstruct · Scale · Retarget · Train · Deliver
  1. 01
    Collect
    Onshore capture — video, teleop, hand demos.
  2. 02
    Reconstruct
    Real workcell → URDF + MJCF + USDA digital twin.
  3. 03
    Scale
    10× via jitter, time-stretch, pose-offset, Blender.
  4. 04
    Retarget
    Remapped across Franka, SO-100, Allegro and four more.
  5. 05
    Train
    LoRA on OpenVLA, π0, RT-X, Octo, RDT, GR00T N1.7.
  6. 06
    Deliver
    tar.gz · dataset card · SHA-256 audit chain.

Stage outputs and per-stage backend modes are visible inside every delivery's Honesty Matrix tab — REAL, MOCK (deterministic stub), or FIXTURE (bundled sample) per stage, with no hand-waving.

Cross-Embodiment Coverage

One dataset.
Seven robots.

One Lane-1 demonstration retargets across the full gripper matrix — parallel-jaw to dexterous, 1 DoF to 24 DoF. Customers pay once, deploy across the fleet.

Franka Panda
2 DoF · PARALLEL
Closed-form exact IK · industrial reference
REAL
Robotiq 2F-85
2 DoF · PARALLEL
Closed-form exact IK · cobot standard
REAL
Robotiq 2F-140
2 DoF · PARALLEL
Closed-form exact IK · wider span
REAL
SO-100
1 DoF · PARALLEL
Closed-form exact IK · low-cost teaching
REAL
Allegro Hand
16 DoF · DEXTEROUS
Real per-finger IK via ikpy · simplified chain
REAL
Shadow Hand
24 DoF · DEXTEROUS
Best-effort fallback today · full URDF on roadmap
PARTIAL
AnyTeach
7 DoF · DEXTEROUS
Best-effort fallback today · full URDF on roadmap
PARTIAL
+ Custom
Bring your own URDF
Roadmap: Inspire Hand, LEAP Hand, Galbot G1 hand — propose your fleet
ROADMAP

Quality + physical-consistency gated by ROPA framework (ICRA 2026) — contact / kinematic / rejection metrics per retarget. NaN propagation from occluded source frames is guarded; corrupt input does not silently corrupt downstream training data.

NVIDIA Full-Stack

Full NVIDIA stack integration.
Without the lock-in.

Three NVIDIA layers wired end-to-end through our pipeline — GR00T N1.7 for policy training, Cosmos Predict 2.5 for generative simulation, Isaac Sim 5.0 OpenUSD for runtime. Non-NVIDIA backbones (π0 · OpenVLA · RDT) equally first-class.

LAYER 01
GR00T N1.7
Policy Training Backbone

2 B-param Vision-Language-Action foundation model. LoRA finetuning over our PIECE 2 trajectory bundles. Customer-selectable in our training pipeline.

Status Live · production-wired
LAYER 02
Cosmos Predict 2.5
Generative Simulation

Text/image/video → physics-consistent simulation scenes. Synthetic augmentation amplifier on top of captive real teleop — 1:10 expansion law.

Status Live · synth-gen backend
LAYER 03
Isaac Sim 5.0
OpenUSD Runtime

5-robot OpenUSD schema (Franka · SO-100 · Galbot · Unitree · Tesla Optimus). PIECE 1 emits USDA directly — drop-in to Isaac without conversion.

Status Live · ships in PIECE 1
Same PIECE 1 deliverable. Any sim stack. Genesis · MuJoCo · SAPIEN customers ship from the same pipeline as Isaac Sim users. No NVIDIA dependency required.
Discuss your stack →
ROPA Physics Validation

Augmentation doesn't break
kinematic constraints.

Three quantified metrics, every augment + retarget run. Threshold gating — caveats logged to PIPL audit chain when partial, batches under fail threshold are not delivered.

Doesn't your augmentation break the kinematics? — what every serious robotics customer asks during due-diligence. We answer with three numbers, not a demo video.
METRIC 01
Contact Constraint Satisfaction
grasp closure validity (higher better)

Gripper-to-object contact normal alignment + grasp closure validity. Computed per trial in physics replay.

Pass≥ 90%
Caveat75-90%
Fail< 75%
METRIC 02
Kinematic Validity
joint-limit compliance (higher better)

Per-frame joint-limit / velocity / acceleration compliance against target robot URDF.

Pass≥ 95%
Caveat85-95%
Fail< 85%
METRIC 03
Rejection Rate
dropped trial fraction (lower better)

Trials failing either contact OR kinematic check, dropped from delivered batch. Counted against bundled trial count.

Pass≤ 5%
Caveat5-15%
Fail> 15%

Reference framework: ROPA · Robotic Optimization Procedural Augmentation · ICRA 2026 · arxiv:2509.19454

PIPL Audit Chain

SHA-256 hash chain.
Third-party verifiable.

Every operator action, every stage output, every retarget event signed and hash-chained. Customer regulators can re-verify with one shell command — no Teehoo access needed.

EVENT 0001
teleop_start
a3f7…
EVENT 0002
stage_extract
9c12…
EVENT 0003
stage_augment
5d8e…
EVENT 0004
bundle_delivered
f04b…
event_id
UUID
timestamp
ISO 8601
stage
enum
actor
operator_id
artifact_uri
string
artifact_sha
SHA-256
backend_mode
REAL/MOCK
caveats
[]string
pii_class
PIPL tier
prev_event_hash
SHA-256
self_hash
SHA-256
+ 11 fields
total schema
# Customer-side verification — no Teehoo access needed
$ cat audit.jsonl | jq -r '.self_hash' > hashes.txt
$ python verify_chain.py audit.jsonl
✓ 1,847 events · chain intact · SHA-256 head: f04b8d…
Lane 1 · Captive Ops

Captive teleoperation,
onshore by design.

Two self-operated bases in Anhui + Jiangxi — not a brokered vendor pool. Data never leaves PRC at capture; PIPL § 38 cross-border path runs separately and audit-gated.

PRC · ONSHORE
Captive Lane 1 · 2 cities · co-located QA
Anhui · Anqing
PRIMARY · CAPTIVE
200 ops · Q3 ramp
Jiangxi · Nanchang
SECONDARY · BROKERED
300 ops · spec-bid
Active Operators
200+
Captive Anhui base · onshore, salaried, PIPL-cleared, ramping toward 1,500 at full production scale.
Brokered Capacity
300+
Jiangxi spec-bid partners · burst capacity for SOW-anchored campaigns.
Production-Scale Target
1,500+
Order-driven ramp — not plan-driven. Hire signal gated by SOW pipeline ≥ 3× current throughput.
Why captive, not full broker. A captive Anhui base — single legal entity, single audit chain, single QA contract — is the only structure that survives both customer due-diligence and PIPL regulator review. Brokered Jiangxi capacity adds elasticity without adding compliance perimeter.
What We Broker

Five data archetypes,
one routing layer.

From real-world capture sites to synthetic simulation, Teehoo brokers a curated pool of vetted capacity partners across China and APAC — exposed to global AI labs through a single contract.

Real-World Capture

Industrial, residential, automotive, retail, hospitality — full-scenario data collection sites with rapid 14-day activation SOP.

VR · MoCap · First-person

Teleoperation

Humanoid, dexterous, mobile manipulation — multi-modal capture (RGB-D, 6DoF, force-torque) at sub-mm precision.

Apple VP · Quest · FastUMI

RLHF + Eval

PhD-tier expert cohort scoring for reasoning, agent trajectory review, sim-to-real verification — anchored by rubric calibration.

Tier-gated · Rubric QA

Synthetic + Sim

USD-format physical assets, sim-to-real pipelines compatible with NVIDIA Cosmos and Isaac Lab — bridge real-world gaps cost-effectively.

USD · Isaac Lab · Cosmos

Labeling at Scale

Standardized 2D/3D perception annotation, BEV, LiDAR point cloud, semantic segmentation — production-grade SLA.

MCAP · LeRobot · OXE
Capacity Network

Vetted supply,
orchestrated as one.

Capacity contributed by certified partners across China and APAC. Every partner audited — operational SLA, data security posture, compliance attestation — before joining the pool.

3,000+
Network Operators
10+
Cities Across China
500K+
Real-world Scenes
14d
Deployment SLA
Partner A

Real-World Rapid Deployment

Standardized site activation SOP from blank floor to first production batch in 14 days. Specializes in residential VR first-person capture, industrial scene collection, and automotive cabin data.

3,000+ operators10,000+ m²7 cities14-day activation98% VR fidelity
Engagement profile Long-standing engagements with major tech platforms, automotive OEMs, and humanoid robotics programs across China. Named references available under NDA.
Partner B

Teleoperation Hardware + Capture

Vendor of teleoperation hardware systems for humanoid and dexterous manipulation, with companion data collection service. Includes high-precision tracking, force-torque sensing, and trajectory replay.

1-3mm precisionFastUMI-classMulti-arm6DoF + F/T
Engagement profile Engaged by global industrial OEMs and frontier humanoid programs requiring sub-millimeter precision teleop data. Named references available under NDA.
Partner C

Standardized Labeling Platform

Production-scale 2D/3D perception annotation, BEV data labeling, LiDAR point cloud, lane and semantic segmentation. Self-developed annotation management platform, multi-language support, ISO-aligned QA workflow.

1,000+ enterprise clients2D/3D · BEV · LiDARMulti-language
Engagement profile Multi-year programs with global automakers, Tier-1 suppliers, and frontier model labs. Named references available under NDA.
Partner D

Synthetic Data + Simulation

USD-format physical assets, sim-ready scene library, and synthetic data pipelines compatible with NVIDIA Cosmos and Isaac Lab. Bridges the sim-to-real gap with controllable physically-accurate scenarios.

USD assetsCosmos compatibleIsaac LabSim-to-real ready
Engagement profile Deployed within leading simulation ecosystems by humanoid OEMs and Tier-1 automotive programs. Named references available under NDA.
Partner E

Egocentric / First-Person Capture

VR + motion-capture rigs producing high-fidelity first-person video for imitation learning. National recruitment network across multiple cities; trained operators delivering 98%+ scene fidelity.

VR + MoCap98% fidelity5+ recruitment hubs
Engagement profile Frontier humanoid programs requiring large-scale imitation-learning corpora. Named references available under NDA.
Partner capacities sourced from publicly disclosed information and partner-furnished collateral. Specific identities disclosed under NDA. Direct verification available on request. New partner candidates evaluated against operational SLA, security posture, and compliance attestation before joining the pool.
How It Works

Spec submitted →
kickoff in 14 days.

One MSA. One PM. One QA dashboard. Five capacity types orchestrated behind the scenes — exposed to you as a single throat to choke.

Step 01

Spec Submit

Data type, volume, geography constraints, customer entity, timeline. Plain English or technical spec — we accept both.

⏱ 30-min discovery call
Step 02

Compliance Route

Routing matrix auto-eliminates incompatible suppliers (PII-sensitive data → never crosses to US labs without CAC approval). Audit log generated.

⏱ Automated · <1 hour
Step 03

Quote in 48hr

Supply pool subset ranked by capacity, price, prior performance. Vendor net price + Teehoo margin shown separately — no hidden fees.

⏱ 48-hour SLA
Step 04

Kickoff & Deliver

Rapid-deployment partners activate in days. Milestone-based escrow. MCAP / LeRobot / OXE format output. QA money-back guarantee.

⏱ Kickoff in 14 days
Compliance Routing Engine

Compliance is the product,
not an afterthought.

A 4×4 routing matrix maps data-origin cohort × customer geography to safe-harbor pathways. Hard-coded red lines. Per-project legal-review checklist. Audit log retained for every engagement.

Routing Matrix · Data Origin × Customer
CN-domestic
lab (L1)
Diaspora /
NA Chinese (L2)
US frontier
lab (L3)
Global
commodity (L4)
CA Diaspora
✓ Green
✓ Green
✓ Green
✓ Green
CB CN-onshore
✓ Green
⚠ Review
⛔ Red
✓ Green
CC SE Asia
✓ Green
✓ Green
✓ Green
✓ Green
CD Latin Am.
○ Capacity TBD
✓ Green
✓ Green
✓ Green
Standard route · pre-mapped pathway Per-engagement legal review Hard red line · structurally not routed

Cross-border pathways pre-mapped

China CAC data-export pathway · EU GDPR / SCC · US BIS export classification · jurisdiction-scoped engagements with audit-ready trails.

Hard red lines, structurally enforced

PII-sensitive China-collected real-world data never routes to US frontier labs without explicit CAC approval. Hard-coded in the matrix — not contractually promised, structurally blocked.

One US-side counterparty

Contracts signed with FULLHOUSE ASSET Management LLC (California). APAC engagements optionally route through a Singapore Pte Ltd. Customer never holds a direct Chinese contracting relationship.

Customer offloads legal homework

Per-project legal-review checklist auto-generated. Audit log exportable. Most customer counsels approve in a single review cycle.

Embodied Data Pilot

De-risk your first engagement.
$25-50K · 4 weeks · money-back.

One specific deliverable, your spec. We carry the supply-side risk because we've already vetted our capacity partners. If our QA delivery doesn't meet your written spec, you get a 100% refund within 7 days.

Price $25K – $50K
Timeline Spec lock → kickoff in 14 days → delivery in 4 weeks
Risk reversal 100% refund if QA fails
Example A 500hr first-person kitchen VR data · MCAP format · action labels
Example B 100hr teleop dexterous manipulation · 6DoF + force-torque
Example C Sim-to-real verification report on your existing policy · 10 robotics scenarios
How to start Reply with (1) data type + volume (2) customer entity (3) target deadline — written scope + quote within 48 hours