01 / 16
Research deep dive · 23 June 2026

Cosmos 3
thinks, imagines
and acts.

An omnimodal world model for Physical AI: language, vision, audio and action in one shared framework.

paper by NVIDIAarXiv:2606.02800v4
COSMOS
3
01 · Plain-English TL;DR

The short version

The problem

Robots need a pile of models.

One model sees, another predicts video, another chooses actions. The hand-offs are slow, brittle and expensive to train.

The idea

Use one omnimodal backbone.

Cosmos 3 jointly models language, images, video, audio and action, then switches behaviour through its input-output token arrangement.

The payoff

A shared world representation.

The same foundation can support reasoning, synthetic data generation, future simulation and robot policy post-training.

Think of it as a model that can describe a scene, imagine what happens next, and connect that imagined future to an action.
Source: paper, pp. 5–6. [1]
01
Section break

Why this matters for Physical AI

The key shift is from composing specialist models to training a shared model of world state, time and intervention.

02 · Why it matters

From model stack
to world backbone

Fragmented pipeline
VLM
see + plan
VLA / WAM
choose action
World model
simulate future

Useful components, but every boundary becomes a coordination problem.

Cosmos 3 framing
Observe
language · image · video · audio
Imagine
future visual / audio state
Act
embodiment-aware action

Shared tokens let the model keep context across modalities and tasks.

1

unified sequence format

5

core modalities named in the paper

3

model scales: Edge, Nano, Super

downstream task configurations

The “∞” is a conceptual shorthand, not a paper metric. Source: paper, pp. 5–6. [1]
03 · Use cases

One backbone,
many jobs

understanding

Multimodal reasoning

Interpret images and video, answer questions, infer scene semantics and reason about embodied tasks.

generation

Synthetic training data

Generate diverse image, video and audio-visual scenes to increase coverage for Physical AI training.

simulation

Forward dynamics

Given an observed state and action, predict plausible future visual states for evaluation or planning.

control

Robot policy

Post-train on target data to jointly predict action and its expected visual consequence.

transfer

Control-guided video

Use edge or depth control video as conditioning and generate corresponding RGB video.

multimodal

Audio-visual worlds

Generate video with synchronised audio, using a shared temporal axis across modalities.

Source: paper, Fig. 1 and Sec. 2.2.2. [1]
02
Section break

The architecture

The trick is not simply “put all modalities in one transformer”. Cosmos 3 separates reasoning and generation while keeping them jointly conditioned.

04 · Deep dive

Two token worlds
meet in attention

Packed sequenceAR: text + ViT visionDM: noisy video + audio + action tokensReasoner towercausal self-attentionnext-token predictionAR tokens stay self-containedGenerator towerfull bidirectional attentioniterative denoisingDM attends to AR + DM contextshared joint attentionone backbone · two parameter pathways · shared context
Reasoner

Language and understanding tokens use causal attention, preserving autoregressive integrity.

Generator

Continuous modality tokens use full attention and are denoised, conditioned by the reasoning context.

Adapted from Fig. 5 in the paper. [1]
05 · Deep dive

Action is a first-class modality

Cosmos 3 treats action as a causal variable: it explains how the world changes, not just what the world looks like.

Forward dynamics

Action → future

Given current video and clean actions, denoise the next visual state.

Inverse dynamics

Future → action

Given observed visual transitions, infer the action tokens that explain them.

Policy

Action ↔ future

Jointly predict the intervention and its expected visual consequence.

Unified action interface
Ego pose
camera / vehicle motion
+
Effector pose
hands / end effectors
+
Grasp state
fingertips / open-close
Shared latent action
Source: paper, Fig. 3–4 and Sec. 2.1.3–2.2.2. [1]
06 · Deep dive

Time must mean
the same thing

The problem

Tokens run at different clocks.

Video, audio and action streams have different sampling rates. A one-step token shift cannot automatically mean the same physical duration.

24

video FPS base

25

audio tokens / sec

10–30

action FPS envelope

The solution

Absolute temporal modulation.

Cosmos 3 scales temporal position increments by the stream’s temporal resolution, aligning modalities on a shared physical axis.

δt = TPSbase / TPS 24 FPS video → 1.0 temporal step 16 FPS video → 1.5 temporal step 30 FPS video → 0.8 temporal step
Design implication: multimodal world models need an explicit notion of physical time, not only sequence position.
Source: paper, Sec. 2.4 and Fig. 6. [1]
03
Section break

Data, training
and evidence

The architecture is only half the story. The paper’s strategy is to build a data engine, then specialise the shared model without changing its bones.

07 · Training loop

Pre-train broadly.
Specialise late.

01

Reasoner pre-training

language + visual understanding

02

Generator pre-training

multimodal denoising

03

Mid-training

joint physical-action data

04

Post-training

text-to-image, video, policy

Synthetic data

Five named families

PhyxSim, RobotSim, DriveSim, SynHuman and Warehouse datasets broaden physical interactions and environments.

Structured captions

JSON as control surface

Captions encode layout, lighting, actions, camera motion, state changes and audio cues.

Prompt upsampling

Intent → scene programme

A reasoner expands a short request into a schema-constrained multimodal specification.

Source: paper, Sec. 3–4, Sec. 6.3, Appendix A–C. [1]
08 · Evidence

What the paper reports

The table below reproduces selected values from the paper’s overview. These are the authors’ benchmark results, not an independent reproduction.

Reasoning scores
General73.7
Smart infrastructure62.6
Driving79.3
Selected generation / action values
91.36

Text-to-image*

82.8

Image-to-video

39.7

Robot policy*

* denotes a post-trained Cosmos 3 variant in the paper’s table.

Reported comparative result

In robot forward dynamics, Cosmos3-Nano and Super post-trained from the mid-trained checkpoint reached 25.52 dB and 26.04 dB PSNR, compared with 22.99 dB for Ctrl-World.

The result supports the value of unified action mid-training for downstream adaptation. It doesn’t establish universal superiority across all environments.

Source: paper, Table 1 and Sec. 6.2.5. [1]
09 · Evidence boundaries

What it proves —
and what it doesn’t

paper claim

Broad capability coverage

Cosmos 3 supports a wide range of reasoning, generation, simulation and action modes under one framework.

reasonable inference

Shared representations may reduce hand-off friction

This follows from the architecture, but the paper doesn’t provide a full systems-level cost or reliability comparison against every modular stack.

open question

Real-world closed-loop robustness

The report shows benchmark and post-training results. Safe deployment still needs task-specific evaluation, calibration and failure handling in the target environment.

For Physical AI, a pretty rollout is not enough. The validation target is whether predicted futures and actions remain useful under distribution shift, partial observability and intervention.
Evidence framing based on the paper’s reported experiments and conclusion. [1]
04
Section break

Implementation guide

The practical lesson is to treat Cosmos 3 as a foundation plus an evaluation and data programme, not as a drop-in robot brain.

10 · Step-by-step

A sensible adoption path

01 · Define the loop

State → action → outcome

Choose one bounded workflow. Specify the observation stream, action interface, prediction horizon and success metric.

02 · Start with simulation

Build a replayable testbed

Use logged trajectories or synthetic scenes. Keep seeds, prompts, model versions and environment settings inspectable.

03 · Make time explicit

Align clocks and coordinate frames

Record FPS, action frequency, temporal compression, sensor latency and coordinate conventions. Don’t hide these in preprocessing.

04 · Post-train narrowly

Specialise without forking the world model

Fine-tune on the target embodiment and task, while preserving a shared representation and a held-out generalisation set.

05 · Evaluate counterfactuals

Ask “what happens if?”

Test action consistency, temporal coherence, state changes, object identity, safety constraints and recovery after mistakes.

06 · Gate deployment

Close the loop slowly

Move from offline replay to simulation, shadow mode, low-risk interventions and supervised deployment with rollback.

Implementation recommendations are synthesis, informed by the paper’s architecture, training and evaluation sections. [1]
11 · Takeaways + sources

The useful mental model

Cosmos 3 is a bridge between synthetic worlds and real-world agents.

Its distinctive bet is that reasoning, generation and action become more useful when trained as related views of the same evolving world.

If you’re building with it:

Primary sources
[1] Cosmos 3: Omnimodal World Models for Physical AI · arXiv:2606.02800v4NVIDIA Cosmos code repositoryCosmos 3 model collectionCosmos 3 project page

Deck note: all benchmark figures and technical descriptions are attributed to the supplied paper. Inference and implementation advice are labelled as such.

Built as a self-contained HTML deck. Use ← / →, the buttons, or scroll. Press Home / End to jump.