Anaxi Labs
The human-aware data, model and evaluation engine for physical AI
The scaling paradigm for robot learning. From real work to trainable physics.
We turn real work into true-scale 4D worlds, training signals robot models learn from directly, and the tests that prove them, all at scale.
CMU Robotics Institute
founding team
31 countries
20,000+ contributors
CMU, LMSYS, Arm
partners
Executive summary · five problems, one pipeline
Our mission: train robots at scale, in real places
PROBLEM
Robot data comes from a few lab rooms
PROBLEM
Lab and teleop rigs are costly, slow to scale
PROBLEM
Sim scenes are hand-built, not real places
PROBLEM
Current human video has no robot actions to learn from
PROBLEM
Robots train and test in empty lab scenes, not real places with people
COMMUNITY
Real tasks, in the wild
20,000+ people in 31 countries
01 · CAPTURE
Phones, at low cost
Real scenes, ready to rebuild
02 · CURATE
True-scale 3D/4D assets
Build sim scenes and ground HAPI
03 · HAPI MODEL
Trainable interaction
Trains LLMs, world and robot models
04 · ARENA
Train and test, anywhere
Scored cases and the Robot Launchpad
WE SELL
3D/4D sim scenes · Trainable interaction models · Evaluation · Raw and refined data · Robot marketplace
WHO BUYS
Robot model, world-model and digital-twin developers · LLM labs adding spatial skills · Robot hardware companies · The public
THE BOTTLENECK
Footage of people at work is getting cheap. How they actually work is not: true-scale worlds with every human interaction, for robots to learn from and be tested in, at scale.
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
02
Why now · buyers already pay for all three
Data, world models and evaluation are billion-dollar bets
01 · DATA
$14.3B
Meta for 49% of Scale AI, valuing it at $29B · Jun 2025
$870M → ~$2B
Scale AI revenue, 2024 → 2025 forecast
$2B+
Mercor’s gross run rate selling expert data to AI labs · Jun 2026
02 · WORLD MODELS AND TWINS
20M hrs
of real-world video trained NVIDIA’s Cosmos world models
~$8.2B
AMD’s deal to buy World Labs, a world-model lab · Sep 2026
$21B → $150B
digital-twin market, 2025 → 2030 forecast
03 · EVALUATION
$30M → $100M
Arena’s run rate, within eight months of launch
$3.1B
Arena’s valuation after a $200M Series B · Oct 2026
Same incubator
LMSYS, which incubated Arena, now incubates our robot leaderboard
OUR BET
Physical AI needs all three, built from real work. We rebuild it as true-scale physics worlds and sell all three as one pipeline.
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
03
Team · large-model scaling and robot control
We scaled large models and robots. Now we scale robot learning.
Kate Shen
Product and Operations
- Carnegie Mellon, BS in CS
- PM at Microsoft and Hulu
- Ran two markets for Xiaomi Global (>$20M monthly revenue)
Chiyu Dong, PhD
AI and Robotics Data
- Carnegie Mellon, PhD in Robotics
- Research Fellow, CMU Robotics Institute
- Tech lead, DiDi autonomous driving (team of 50)
Yupeng Liao
Infra
- Principal Architect at Disney+
- Infra for 200M users and PB-scale video
- Tsinghua University, BS in EE
Yuqing Tang, PhD
Large-Model Scaling, World-Model and Evaluation
- Scientist at CMU Robotics Institute, Meta AI, Amazon Alexa AI and Microsoft Research
- Top-5 contributor to fairseq (mBART, wav2vec 2.0)
- Director of XPeng Robotics
Top robotics school and industry network
PB-scale data through real robotics pipelines
Foundation-model training at scale
Global supply chain for data capture
PB-scale video processing
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
04
Partnerships · research, arena, industry standard
Partners: Carnegie Mellon, LMSYS and Arm
RESEARCH
Data evaluation and pricing
- Led by CMU AI professor
- CMU PhDs as researchers; joint papers
INCUBATOR OF
Chatbot Arena
SGLang
EVALUATION ARENA
Incubating our public leaderboard
- Incubated Chatbot Arena, now Arena at $3.1B, and SGLang
- Co-incubates the public board only; our commercial business is ours alone
INDUSTRY STANDARD
Launch partner, Arm Robotics Capability Framework
- Arm powers the chips in most robots
- Member of Arm’s robotics working group, shaping how the industry rates robot capability
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
05
The community · real-world coverage, in the wild
Real tasks, filmed in the wild, across the whole economy
20,000+
contributors
31
countries on 5 continents
Each dot is a real collection point · compliant in every jurisdiction
UN ISIC DIVISIONS
14 of 24
manufacturing divisions on film
74.6% of manufacturing workers are in industries we film
REAL, NOT STAGED
Real clutter, real objects, and people working together
Homes
Warehouses
Factories
Fields
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
06
01 · Capture · phone-based 4D capture
Real-world 4D capture that scales to every corner of the world
Our phone-based 4D capture technology syncs ordinary phones into video, depth and motion on one clock, doing the job of 3D scanners and capture rigs at a fraction of the cost.
THE APP ON 3 PHONES
ONE SYNCED TAKE
01 · LOW COST
Phones people already own
No rigs: film any real scene, anywhere
02 · FULL SCENE
Video, depth and motion
Synced views capture geometry, motion and contact
03 · COMPLETED
Community + agent models
Add physics: size, mass, material, friction
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
07
02 · Curate · for training, scene building and evaluation
Real video becomes 3D/4D assets and HAPI models
Phone recordings become true-scale objects with real dynamics, and thousands of training and test cases.
INPUT · PHONE RECORDINGS
Synced views, depth and motion from the phone app
agents + community
SIM RUN
A robot runs the rebuilt task in Isaac Sim
builds the sim scene
3D/4D ASSETS
Scene building: real-scale objects and their dynamics
grounds the HAPI model
HAPI MODEL
Differentiable geometry, motion, contact and physics
TRAINS
Multimodal LLMs
World models
Robot models
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
08
03 · HAPI model · human-aware physical interaction
HAPI turns real video into trainable interaction models
INTERACTION LAYERS
Geometry · Motion · Contact · Physics
Not just how a world looks: how people handle objects, and how they respond to each other and to robots
01 · GROUND TRUTH
Reconstruct what happened
All four layers, recovered frame by frame from phone video
02 · TRAIN AND TEST
Generate plausible variants
Small shifts in pose, timing and force: fresh episodes to learn from, fresh Arena cases to score
03 · LEARNING SIGNAL
Return precise gradients that train the robot model directly
Not just a sim to try things in: every layer is differentiable, so models see exactly what to fix
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
09
04 · Arena · evaluate robots and robot models
Know a robot works before it ships
Every training video pack is also a versioned case that scores robots and robot models, in sim or on real robots. Every failure orders new footage.
01 · ROBOT RUN
Pick any robot model and any robot; run the curated case
02 · RUN EVIDENCE
Scores, logs, video and human review; failures become data briefs
Test across thousands of real scenarios, not a handful
TODAY
WITH ANAXI
Real scenarios to test in
Dozens
→
Thousands
Time to add 1,000 new ones
Years
→
Weeks
Cost to add 1,000 new ones
$5M+
→
~$40k
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
10
04 · Arena · Robot Launchpad
Try and play with any robot in your place, then deploy it
01 · SCAN YOUR PLACE
Scan any home, office or field into a true-size sim scene
02 · ASK, WATCH, RATE
Ask for anything in chat, watch the run, rate any moment
WHERE IT GOES
CUSTOMER POOL · OPEN-ENDED
A marketplace for everyone
Try, compare and deploy robots and models from any maker
4.7B
mobile internet users can scan and play
Every place
homes, offices, fields, warehouses
$40B
consumer robot spend by 2030
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
11
Business model and milestones · 2026 to 2028
Data lands the account; HAPI and evaluation grow it
Q4 2026 · NOW
1 · LAND
Paid pilot: raw data, filmed to order
PAYS
Base fee + per task-hour
VOLUME
1K → 100K task-hours
2027
2 · EXPAND
Data, HAPI signals and evaluation
PAYS
Per-hour data, signal API, evals
VOLUME
1M task-hours ≈ $63M
2028
3 · RECUR
Regression plans, marketplace
PAYS
Subscriptions, re-licenses
VOLUME
10M+ task-hours ≈ $335M
TOP OF THE FUNNEL
Free leaderboard (incubated by LMSYS) and the Robot Launchpad; payment never buys rank
Each tranche is bought only after Arena shows a gain. $ = 3-year cohort price; 40% margin floor (estimates).
CUSTOMERS
Robot model developers
World-model and digital-twin teams
LLM labs
Robot hardware companies
The public, via the Launchpad
Anaxi Labs
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
→
Business
12
OUR MISSION
Take robot learning out of the lab: real human work in 4D, with its geometry, motion, contact and physics as training signals, at scale.
Community
→
01 Capture
→
02 Curate
→
03 HAPI
→
04 Arena
Thank you
THE SCALING PARADIGM FOR ROBOT LEARNING
Appendix · how robots are simulated today
Simulation is well funded; its worlds are mostly generated
World-model labs raised ~$3.5B in 2026. We rebuild real places in true-scale 4D and turn real work into training signals.
WORLD-MODEL LABS
SIM PLATFORMS AND DATA
Anaxi Labs
World Labs
AMD, ~$8.2B
General Intuition
$674M raised
NVIDIA
Cosmos, Isaac
Moonlake AI
$28M seed
Lightwheel
~$290M raised
Human-video suppliers
Several, funded
Sim worlds built from
Real places, filmed on phones
Text or photo prompts
Video games
Physics sim, synthetic
Photos, scans and CAD
Sim-asset library
None: footage only
Real people doing real work
20,000+ people, 31 countries
—
No: gamers at play
Yes, first-person video
Own team, a few teleop runs
Operators in headsets
Yes, at scale
Real work as trainable physics
Geometry, motion, contact, physics
Generated geometry only
Game actions only
Wrist and hand motion only
Physics tuned by robot teleop
Pose capture only
Footage and pose only
Sold as
Sim scenes, HAPI, evaluation, data
3D world generator
Agent, world models
Platform, open models
Deployment eval
Assets, data, sim eval
Video and labels
Founding team
CMU robotics PhD, LLMs, robot industry
Vision and graphics researchers
Game-app founders, AI researchers
Large robotics research team
Stanford AI PhDs, sim and 3D
NVIDIA and Cruise sim lead
Varies
Anaxi Labs
A1
Appendix · how robots are tested today
Robot testing is early: each test covers one setting
Paid robot evaluation is only starting, and no one tests across real workplaces at scale.
ACADEMIC · MOST CITED
COMMERCIAL · LAUNCHED 2025–26
Anaxi Labs
RoboCasa365
LIBERO
RoboDojo
AA-Robotics
PhAIL
Moonlake AI
Scale AI · Lightwheel
Work domains covered
17 domains
1 · kitchens
1 · homes
1 · tabletop
1 · tabletop
1 · tabletop
Per customer site
Not published
Where the scenes come from
Real workplaces, rebuilt
Simulated kitchens
Simulated homes
Sim + lab arms
One lab arm
One lab arm
Sim from photos, CAD
Closed
Tasks in the test
231, growing at scale
365, fixed
130, fixed and saturated
60, fixed
8
144, fixed
Built one task at a time
Not published
Built as
Industrial-scale product
Academic project
Academic project
Academic consortium
Benchmark firm
Start-up
Start-up
Data vendors
Turns failures into data briefs
Yes, down to the object
No
No
No
No
No
Model tuning, no data
Sells data, no neutral test
Survey of 47 robot evaluators, Oct 2026. RoboTwin 2.0 and real-robot boards such as RoboArena, Robocurve and RoboChallenge also test tabletop tasks.
Anaxi Labs
A2
Appendix · the performance matrix
Which model, on which robot, is ready for which job
JOBS (GDP SECTOR · TASK)
MODEL ON ROBOT →
Model A
Arm X
Model A
Humanoid Y
Model B
Arm X
Model B
Mobile Z
Model C
Humanoid Y
Manufacturing
pack cartons
92
81
86
64
58
Manufacturing
fasten screws
71
44
83
39
48
Retail
restock shelves
38
46
42
35
41
Food service
plate meals
88
62
74
57
85
Logistics
sort parcels
90
84
87
91
66
Construction
drill anchors
47
31
52
28
36
Healthcare
deliver supplies
55
49
61
89
72
80+ passes the bar
50–79 close
under 50 falls short
Each cell is tagged sim, real site, or both.
READ ACROSS → WHO DEPLOYS WHAT
Buyers: who is ready
Four pairs pass for sorting parcels. Payment never buys rank.
Pass → ready to deploy for that job
READ DOWN → WHAT TO FIX
Developers: what to fix next
Fails on every pair: a skill gap that becomes a data brief
Retail › restock › open › drawer
Fails on one robot only: fix the robot, not the model
Anaxi Labs
A3