Ontos

Agentic being in the world.

{ }

coding

∑

mathematics

Aa

language

Superhuman.

Text reasoning works.

The bet that worked

METR · metr.org/time-horizons

The world is multimodal.

pixels · motion · contact · a body moving through a room

Text reasoning works.?

The demo everyone is chasing

"Agentic AI looks ready for the physical world."

Waddle Labs, July 2026

Real hardware, a competitor's footage, attributed.

But off their training distribution, none of today's models generalize.

Anthropic Frontier Red Team, 9 July 2026

No Anthropic model successfully stood the robot up from a collapsed pose even once.

Humanoid under model control · Anthropic

0

humanoid stand-ups

5.5%

best manipulation success

~0.4

best score, out of 1.0

The world is multimodal.
Yet reasoning is still largely textual.

Pixels, forces, states. Text reasoning has nothing to perceive any of it with, and nothing to check itself against. That gap is physical AI.

Two bets for Physical AI: teleoperation and world models.

Teleoperation

Expensive. Not generalizable.

World models

We can scrape the internet for videos, but how about actions?

[Google DeepMind] Genie 3

Teleoperation:

1. Expensive: One hour of robot data costs more than an hour of human labor.

$120/h

fully loaded 2026 market rate: operator, rig depreciation, QA

2026 market benchmark

$50–150k

the data bill for one new task: 300 to 1,200 demonstrations

2026 market benchmark

2–3h

usable data from one full 8-hour operator shift

operator-shift yield

2. Poor generalization ability: teleoperation + behavior cloning.

Trained on: blue

Tested on: red

[Google DeepMind] ALOHA Unleashed: A Simple Recipe for Robot Dexterity. Trained to fold a blue shirt, it generalizes only to a blue shirt.

SAY

Robot data has a price and it is public. About a hundred and twenty dollars an hour, fully loaded: the operator, the rig depreciation, the quality control. That is the 2026 market rate. And it has come down, to be fair. It was three hundred and forty dollars an hour in 2024.

Here is what that buys. Training one new task takes three hundred to twelve hundred demonstrations, so the data bill for a single task lands between fifty and a hundred and fifty thousand dollars. One task. Not a product, not a robot that clears your table. One task.

The third number is why this never gets better. A trained operator working a full eight-hour shift yields two to three hours of usable data. The rest goes to setup, calibration drift, failed takes, sync errors. One hour of robot data costs roughly one hour of skilled human labor, and you cannot speed up a demonstration any more than you can speed up the robot.

So the cost per hour has a floor, and the floor is a person. Text got cheaper as it got bigger because the next token was free. Here the next hour costs a hundred and twenty dollars, forever. If the road to general robotics runs through teleoperation, the bill scales linearly with capability and nobody can pay it. That is the assumption we think is wrong, and what we build next is the alternative.

BACKUP for Q&A

The three headline numbers are market benchmarks, not peer-reviewed results, so say "market rate" and not "study" if pressed. The $118 to $200 per hour range is corroborated independently by VDC Strategy ("roughly $120 per operator-hour as of early 2026, down from around $340/hour two years earlier"), the aicost.ai teleoperation calculator ("$118-200/hr, was $340 in 2024"), NeuralChainAI ("$118 to $200 per collected hour") and DataX Power ("$80-150 per hour" for full humanoid multi-sensor programs). The 300 to 1,200 demonstrations per task and the resulting $50k to $150k per-task budget come from the same set. The 8-hour shift yielding 2 to 3 usable hours is NeuralChainAI, which publishes the yield math: the loss goes to setup, calibration drift, failed takes, consent gaps and sync errors. VDC Strategy has the cleanest one-line version of the headline: "one hour of robot training data still requires roughly one hour of skilled human operation, because you cannot speed up the demonstrations any more than you can speed up the robot."

DROID (Khazatsky et al., RSS 2024): "76k demonstration trajectories or 350 hours of interaction data, collected across 564 scenes and 86 tasks by 50 data collectors" on 18 Franka Panda arms at 13 institutions over 12 months. Note those are successful trajectories only, which is exactly the gap PEVA addresses.

Mobile ALOHA (CoRL 2024): "we build Mobile ALOHA with a $32k budget, comparable to a single industrial cobot such as the Franka Emika Panda," including onboard power and compute. The same paper notes its demos came from two expert operators and flags learning from suboptimal, heterogeneous data as future work.

Scale AI Physical AI page, live: "1000+ hours of diverse demonstration data are collected and uploaded to Scale's data platform every day." Their Sept 2025 blog claims "more than 100,000 production hours" to date and names Physical Intelligence, Generalist AI and Cobot as customers.

Expect the synthetic data objection, because it is the obvious one. 2026 reports do claim synthetic augmentation can offset up to 40% of real demonstrations. The answer is that synthetic generation amplifies a real corpus, it cannot originate one: you cannot simulate contact dynamics, friction coefficients or sensor noise you have never observed, so every synthetic variant still traces back to a real seed hour that somebody paid $120 for. Note NVIDIA reached the same conclusion at GTC 2026, naming demonstration data the bottleneck for physical AI even while projecting heavy synthetic augmentation downstream.

If someone wants peer-reviewed numbers instead of market rates, fall back to the two citations above: a $32k rig from Mobile ALOHA at CoRL 2024, and DROID at RSS 2024 producing 350 hours from 18 robots and 50 operators across 13 institutions in 12 months, which is roughly 19 hours per robot per year. Those are the floor and they tell the same story. One caution on a number in circulation: the "$50k to $150k" figure is widely misquoted as the cost of a single rig, including in an SCSP guest op-ed, but the industry sources use that range for the per-task data budget. Use it the way this slide uses it. Removed the $55.8B Dealroom VC figure entirely: H1 only, Dealroom's broadest methodology, defense-skewed, and not verifiable from Dealroom directly. This slide's payoff lands on 28, the interaction data flywheel, and slide 33 reuses this cost basis directly.

Current World Models:

1. Synthetic data: action space is different, hard to transfer to real life.

DreamerV3 — one world model, many simulated domains, none of them the real one.

2. Real video with actions: still need collection.

Ma et al., "Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild"

AGPI · Artificial General Physical Intelligence

The fastest path to AGPI fuses World Models with Agentic Computing: Recursive Agentic World Models.

Generalizable. Works across all domains. Superhuman.

Recursive Agentic World Models

What no one else has

Two levels at once

Predicts low-level latent appearance streams and high-level affordance trajectories in one model.

No action labels required

4D motion read out of unlabelled 2D video is the action.

Reliable reaction

A frame only teaches the model something when perception reads it right.

Agentic deployment as the data flywheel

The agentic layer picks what to attempt, so data collects itself in the field.

Ontos One

Recursive World Models

Predicting pixels, latents, and 4D trajectories, all self-trained.

[Nexar] BADAS · Real World Collision Anticipation

BADAS: Real-World Latent World Models

Pedestrian

Forklift

Collision

Sidewalk delivery

Aerial drone

PEVA · Whole-Body Conditioned Egocentric Video Prediction

A World Model for Embodied Agents: the first whole body conditioned world model

What would a World Model look like if we start from a real embodied agent acting in the real world? It has to have:

A real, physically grounded and complex action space, not just abstract control signals.
Diverse, real-life scenarios and activities.

Past Frames

whole-body action as a relative 3D pose delta

Next Whole-Body Action

Predictions

Embodiment-agnostic

Actions are relative 3D pose deltas, independent of any specific hardware or control system.

Not just success

Learns from failure and exploration equally, not only expert trajectories.

What-if in pixels

Given history frames and an action, anticipate the consequence. Change the action, the future changes.

PEVA · planning

Planning in the real world by simulating counterfactuals

Trained on real egocentric video and full-body motion capture, not synthetic physics engines. Sample actions → roll them out → keep the path closest to the goal across complex, uncontrolled environments.

Input

Goal

World Motion Models · Flexible Sequence Modeling of SE(3) Trajectory

4D World Motion Model Enables Strong Generalization

Representation · video to SE(3) trajectory

Method · TraceGen

RGB · query points

lifted to 3D

Input · one frame

TraceGen

2D tracks

3D tracks

Generated · T = 32

Application · one model, every domain

Recursive world model

A recursive world model
that self-improves

Predicting pixels, latents, and 4D trajectories, all self-trained.

Video

unlabelled · internet scale

Structured latent

▲ 4D motion ◆ geometry ◈ objects

4D motion is a_t. No labels needed

World model

o_t + a_t → o_t+1

predict the next frames

New observations

real frames come back

Act in the world

execute a_t

self-conditioning every turn is new data

Ontos One

Multi-stream Reaction

Reliable perception and action.

Change-aware multimodal reasoning.

Ontos One knows where to look.

Problem Setting

Five changes. You found them by looking back and forth — the encoder never gets to.

Method

Stateless

Every frame embedded alone.

Stateful (ours)

Carries state across frames — what moved is legible in the embedding itself.

Sample output: "The small blue rubber sphere was moved." Read directly off the embedding, no extra reasoning step.

Did you realize the two images are different?

Five changes.

Two images.

And when you were comparing, did you eyeball the two images back and forth?

Current visual encoders in AI models can't.

BUILD (5 clicks, matches the original SVE deck): lands on the pair, captioned "Two images." — just let them look. Click 1: the caption turns into the question, "did you realize the two images are different?" — let the room hunt. Click 2: five rings, "Five changes." Click 3: rings fade out, pair lifts, "did you eyeball the two images back and forth?" Click 4: "Current visual encoders in AI models can't." Click 5: clears to the problem/method slide underneath. Merged from two slides. From the ontos26 pptx (slides 7-12). Problem: current visual encoders bottleneck the LLM's ability to represent change between frames — the two CLEVR-style frames differ by one swapped object, easy for a human to catch. Fix: make the encoder itself change-aware (Stateful Visual Encoder, SVE) — a stateless encoder embeds Image 1 and Image 2 independently and hands both to the LLM to difference; the stateful encoder folds Image 2's embedding through the state left by Image 1.

Change-aware multimodal reasoning.

Impacts on real-world tasks

Image editing

Reference — Replace person located slightly off-center in the upper-middle section with a large rock.

Baseline — Remove the person standing on the rocky outcrop, arms raised, lower-middle of the image.

+ Ontos — Replace person positioned in the upper-central area of the image with a large rock.

Remote sensing

Reference — Some detached houses appear beside the road on the bareland.

Baseline — The scene is the same as before.

+ Ontos — A building appears at the top of the scene.

Medical imaging

Reference — The main image has additional findings of lung opacity and atelectasis vs. the reference image. It is missing the finding of cardiomegaly vs. the reference image.

Baseline — The main image has additional findings of atelectasis and cardiomegaly. It is missing the findings of pleural effusion and lung opacity.

+ Ontos — The main image has additional findings of atelectasis and lung opacity. It is missing the finding of cardiomegaly.

Tactile-Reactive Dexterous Manipulation

Ontos One reacts “fast and slow”.

Slow system

Understand the goal, plan long-horizon actions, use language and vision, predict the future.

Fast system

React to contact, adjust force and trajectory, correct slip and errors, keep contact stable.

~5 Hz

slow system: deliberate reasoning

~20 Hz

fast system: reactive tactile reflex

Ontos augments VLAs with a fast tactile reflex pathway. Slow deliberation sets the direction; fast reaction delivers the dexterity.

Multi-stream Reaction

Ontos One self calibrates.

C³: Calibrated Conformal Control for Verifiably Steerable Foundation Models

Ontos One

Agentic Self-improvement

Agentic harness for Physical AI.

Agentic Self-improvement

Code as Policy

[Google] Code-as-Policies: an LLM turns an instruction into a program built from perception and control primitives. Liang et al., ICRA 2023 · code-as-policies.github.io

RATs · Robotic Agent Teams

A recursive world model, that self-coordinates.

Ontos One, the first Agentic Harness for Physical AI with multimodal reasoning,
via Recursive World Models and Reliable Reaction.

Short-term reasoning

Skills

A multimodal LLM–VLA that runs one module per modality, each at its own rate — vision, tactile, language, sound. Foveation, comparison, grounding.

Mid-term reasoning

Imagination

A recursive world model built on a 4D-conditioned latent sequence model — and the 4D structure is self-trained from 2D pixel prediction, so no one has to label it.

Long-term reasoning

Continual self-improvement

An agentic robotic harness that imagines futures, forms plans, then executes, refines and banks the skill it just learned.

Competitive landscape

No one else has the full stack a physical agent needs.

World modelto plan in

Agenticself-improves

Ontosone stack: reason, plan, improve

Skild · Physical Intelligenceteleoperation → VLA policy

Waddle · OpenAI · Anthropictext models given a body

Ekavision-force-action, learned in sim

World Labs · Genesis · Moonvalleygenerative world models · Genie 3

AMIcognitive architecture, research

SAY: the last twenty minutes named three things. A world model you can plan in, on a real body. Reasoning in the world, not in text. An agentic loop that decides what to learn after it ships. That is the table. Walk it left to right, not company by company. Teleop plus VLA — Skild, Physical Intelligence, and every other policy lab — is the first failed bet from earlier: expensive demonstrations, frozen at deploy. They have vision-language grounding, which is why reasoning is partial; they do not have a world model, they do not keep improving. Be fair when pressed: they ARE cross-embodiment. The difference is what the model outputs. Waddle, OpenAI, Anthropic is the demo everyone is chasing. Text reasoning wrapped onto a body. Agentic in software — Claude Code is real — and empty in the physical world. Anthropic's own eval is the evidence: zero stand-ups. If that demo generalized there would be no company. Eka is the sim-and-force variant. Dexterity is real. The world model is a simulator, so it stays partial. World Labs, Genesis, Moonvalley, Genie 3 — this is the second failed bet. They genuinely have world models. That is their whole product. Mark them full, then say the rest of the row is empty: no reasoning in the world, no agentic loop, and the action space does not transfer onto a real embodiment. AMI is the closest cousin. World models as a cognitive architecture, research-grade reasoning and a research-grade agentic story. No deployment loop. Nobody else is full across the row. That is the only sentence this slide has to land. Next is the market.

Trillion-dollar market for Physical AI

drivable-area segmentation on a city street

BDD100K · CVPR 20

Autonomous vehicles

pixels, depth, motion

ARM4R · ICML 25

Robotics

vision, action, force

overhead satellite view with a segmented target

LISAt · NeurIPS 25

Geospatial & defense

SAR, multi-spectral, volumetric

CT Seg · arXiv 26

Medical imaging

3D volumes, small changes

generated environment rolled out by a world model

NWM · CVPR 25

Gaming & simulation

space and time, interactive

everyday indoor scene from a head-mounted camera

PEVA · NeurIPS 25

Home & office cameras

continuous, everyday scenes

Competitive landscape

Physical AI is not a robotics category

AVvehicles

Roboticsbodies

Geodefense

Simgaming

Cameraseveryday

Ontospublished in all five

Skild · Physical Intelligenceteleoperation → VLA policy

Waddle · OpenAI · Anthropictext models given a body

Ekavision-force-action, learned in sim

Waabi · Wayveend-to-end driving world models

RAIC Labs · Capellaunlabeled GEOINT · SAR

World Labs · Genesis · Moonvalleygenerative world models · Genie 3

AMIcognitive architecture, research

SAY: robotics is where the public failures are measured, so the room files us as a robotics company. This slide stops that. Same idea as the last slide, now five domains: vehicles, robotics, geo and defense, gaming and simulation, everyday cameras. Ontos is full across because we already publish in all five — BADAS, T-Rex, LISAt, neural simulation, PEVA. That is the proof, not a TAM slide. Do not put Nexar on this table. Luc is their Chief R&D Officer, BADAS is their product and our AV receipt, and they sit on the partner slide. They are supply, not a peer. Waabi and Wayve are the honest AV row: Waabi World and Wayve GAIA-4 are real driving world models, shipped. Sim is partial because that simulator is a driving harness, not a general world. Robotics and geo stay empty. RAIC Labs (ex-Synthetaic) and Capella are the honest geo row: unlabeled satellite/ISR search, and a SAR constellation with NRO contracts. Full on geo, empty everywhere else. Skild and PI are a robotics row. Waddle is a robotics demo, hence partial. Eka is robotics. World Labs and Genie actually do own simulation and a slice of video — do not over-claim the gap. AMI is research in robotics and simulation, nothing shipped in the others. Land: specialists exist in every column. Nobody else is full across the row. The company is the method that generalizes. Then four starting impact areas.

Four key impact areas

Two we already showed.

01

Real world robotics

Sample efficient dexterous manipulation learning leading to self-exploration and learning. Carry stacked dishes and cups like human waiter. Use a Sharpa hand to manipulate chopsticks to pick up sushi, or an ice cube.

02

Gaming

PEVA++ enabled cross embodiment mocap transfer: agile and on demand reconfigurable NPCs.

03

Geospatial

Reasoning over diverse, heterogeneous Geospatial data. We need to infer complex relationships across and between sensors to understand environmental risk.

Satellite, drone, camera, acoustic, wildlife, and mobile sensors overlaid on the globe

UAV, satellite, bioacoustic, camera-trap, and animal-sensor streams fused into species, biomass, and occupancy

Spotted lanternfly nymphs on a plant stem

Overhead satellite view of coastline, fields, and settlement

Kahramanmaraş, Turkey — 7 February 2023 — Planet Labs

04

Environment monitoring

Build easily, launch fast, and connect with millions in the new Ring Appstore.

ring Developer

Ring alarm, keypad, doorbell, camera, and sensors

Business Intelligence & Operations

Event-based notifications, workflow automation, customer traffic analysis.

Elderly Care Monitoring

Motion analysis, activity alerts, daily summaries.

Pet Wellness Assessment

Behavior pattern analysis, health insights, mood detection.

Home Analytics

Pool condition monitoring, package theft detection.

Amazon Ring Marketplace API — home and office cameras. Pipeline, not signed.

The build plan

The first twelve months

Months 0–3

One model per layer

Data and GPU deals
Open-source → proprietary: one model per layer
SOTA on at least one domain, each layer
Serve free, rate-limited

Months 3–6

Merge into AGPI

Scale and merge across domains
Scale and merge across layers
Serve an AGPI model, breadth increasing

Months 6–12

Distill, sell, edge

Modularize and distill AGPI
Serve / sell proprietary models
Efficient edge, or hybrid edge/cloud

AGPI — artificial general physical intelligence — is the merged model these three capabilities fuse into.

The team

Luc Vincent

Co-founder & CEO

Google Street View · Lyft Level 5 · Meta · BADAS

Trevor Darrell

Co-founder, President & CSO

Berkeley & MIT professor · founded BAIR · Caffe, FCN, LRCN · Pinterest, Prompt, Voio, SafelyYou, Nexar

Yutong Bai

Co-founder & CTO

LVM · PEVA · Google Brain, Meta, Berkeley

Colin Wang

Founding Engineer

CharXiv · SVE · Princeton, Berkeley

The first ‘World’ Model: Google Street View

The first planet-scale capture system. Built by our CEO.

LVM · CVPR 2024

Large Vision Model: the First Scalable Vision-only World model

We figured out how to reason directly in raw sensory data: no language supervision, no curated labels. That unlocks a structural data advantage others cannot replicate.

Capabilities

Trains on raw, unlabeled visual streams at internet scale
Learns from visual context alone, no text labels required
Data flywheel: every modality, every environment feeds the model

420B

visual tokens, zero language

7B

parameters · one epoch

Visual sentences unify images, video, and annotations into one training format

Visual sentences · UVD-V1 · CVPR 2024

Moat

01

Speed and experience

The team has shipped this class of model before — we get to a working system faster than anyone starting from scratch.

02

Unique architecture

Not something an average team reproduces. The design choices come from having built the prior generation.

03

Open-source and proprietary

Open weights build the ecosystem and the funnel; the proprietary stack stays ours.

04

Partner network

Fleets, labs and design partners already in reach — distribution and interaction data from day one.

Hiring

4

founders

8

research stars

2

infra / MLOps

2

ops

16

total headcount

The eight research hires

Experience spread evenly across the four domains — two in each, so no single vertical carries the model.

Robotics · 2 Gaming · 2 Geospatial · 2 Environmental monitoring · 2

Partner network

Possible data deals

Nexar

LVTD

Driving footage. BADAS already in the stack.

UCB HIC partners

TD

GM, Panasonic, Bosch, Samsung, …

Robot.com

TD

Robotics fleet access.

EA

TD

Gaming and simulation.

Amazon Ring Marketplace API

Home and office cameras.

Pipeline, not signed. Initials are who owns the relationship.

The raise: $100M · room to $180M

Two-year runway. Same plan — the delta is compute.

Base

$100M

~$49M / yr × 2

GPUs$40M/yr

Team$7M/yr

Facilities$1M/yr

Data$1M/yr

Run-rate~$49M/yr

Accelerated

$180M

GPUs scale · team & data grow

GPUs$80M/yr

Team$7M/yr +%

Facilities$1M/yr

Data$1M/yr +%

2-yr capital~$180M

Primary ask $100M at $1B post. $180M is the same milestones with compute bought forward. · AMI · World Labs · Skild · Liquid AI · Physical Intelligence

ONTOS

Reasoning, rebuilt for
the multimodal world.

Backup

Multi-stream Reaction

Ontos has scalable modularity.

FlexOlmo: Open Language Models for Flexible Data Use

Reasoning stands in the way

01

Data

There is no internet of physical interaction.

02

Reasoning

Plan, search, and backtrack still happen in text, not in the world.

03

Real time

The loop closes in milliseconds, not seconds.

Go to market & monetization

The motion The meter

LandMonths 0–3

Design partners who already own fleets and already buy from vendors.

Data for credits — partners pay in interaction data, not cash. Free tier alongside it, rate limited.

ProveMonths 3–6

One embodiment against their baseline, then a second embodiment.

Paid API — usage-based pricing once the model beats that baseline.

ExpandMonths 6–12

Serving across the fleet, deployment data compounding into what we train next.

Enterprise / edge API — dedicated and on-device deployments.

Revenue lines are cumulative — the earlier ones keep running as the next one opens.

To decide: which partners, on what contract, and the date of the first dollar

How we expect the first deals to be structured

Data for credit

Fleet access, paid in model credits.

Data for equity

Sustained access, for a stake.

Compute for equity

Offsets the $40M training line.

Design partner, cash

Paid deployment from day one.

NVIDIA · Panasonic · Bosch · Apple · Robot.com · EA · Kitware / NGA · Ring

To decide: status of each counterparty: live, warm intro, or illustrative

Founding faculty advisors

Sewon Min

UC Berkeley

FlexOlmo

David Chan

UC Irvine

Reverse

Roei Herzig

Simon Fraser University

T-Rex

Ritwik Gupta

University of Maryland

LISAt

Sara Beery

MIT

WILDS

COVT · Chain-of-Visual-Thought

Teaching VLMs to See and Think Better with Continuous Visual Tokens

What it does

COVT interleaves discrete text tokens and continuous visual tokens inside think tags

Result

Counting

Relative depth

Scene understanding

Agentic being in the world.

Superhuman.

Text reasoning works.

Text reasoning works.?

But off their training distribution, none of today's models generalize.

No Anthropic model successfully stood the robot up from a collapsed pose even once.

The world is multimodal.Yet reasoning is still largely textual.

Two bets for Physical AI: teleoperation and world models.

Teleoperation

World models

Teleoperation:

Current World Models:

The fastest path to AGPI fuses World Models with Agentic Computing: Recursive Agentic World Models.

What no one else has

Recursive World Models

BADAS: Real-World Latent World Models

Pedestrian

Forklift

Collision

Sidewalk delivery

Aerial drone

A World Model for Embodied Agents: the first whole body conditioned world model

Planning in the real world by simulating counterfactuals

4D World Motion Model Enables Strong Generalization

A recursive world modelthat self-improves

Multi-stream Reaction

Ontos One knows where to look.

Stateless

Stateful (ours)

Impacts on real-world tasks

Ontos One reacts “fast and slow”.

Slow system

Fast system

Ontos One self calibrates.

Agentic Self-improvement

Code as Policy

A recursive world model, that self-coordinates.

Ontos One, the first Agentic Harness for Physical AI with multimodal reasoning,via Recursive World Models and Reliable Reaction.

Skills

Imagination

Continual self-improvement

No one else has the full stack a physical agent needs.

Trillion-dollar market for Physical AI

Autonomous vehicles

Robotics

Geospatial & defense

Medical imaging

Gaming & simulation

Home & office cameras

Physical AI is not a robotics category

Four key impact areas

Real world robotics

Gaming

Geospatial

Environment monitoring

Business Intelligence & Operations

Elderly Care Monitoring

Pet Wellness Assessment

Home Analytics

The first twelve months

One model per layer

Merge into AGPI

Distill, sell, edge

The team

Luc Vincent

Trevor Darrell

Yutong Bai

Colin Wang

The first ‘World’ Model: Google Street View

Large Vision Model: the First Scalable Vision-only World model

Moat

Speed and experience

Unique architecture

Open-source and proprietary

Partner network

Hiring

Possible data deals

The raise: $100M · room to $180M

Reasoning, rebuilt forthe multimodal world.

Backup

The world is multimodal.
Yet reasoning is still largely textual.

A recursive world model
that self-improves

Ontos One, the first Agentic Harness for Physical AI with multimodal reasoning,
via Recursive World Models and Reliable Reaction.

Reasoning, rebuilt for
the multimodal world.