The world's most popular AI video generation platforms are racing into robotics along with seemingly everyone else in the tech world.
Heading into 2026, popular tools like Google's Veo and Runway are being framed as foundational to the predicted rise of the machines. They're entering a race toward so-called world models, which are AI systems that can stand in for reality itself.
Billions of dollars are flowing into the emerging field as skepticism grows about large language models as an enabling layer in the age of embodied artificial intelligence.
They're basically stepping on NVIDIA's turf, similar to how the GPU maker expanded beyond graphics and into AI computing in the mid 2000s. They're taking systems built to render video and retrofitting them as infrastructure for simulating reality itself.
Since AI video generation took off a few years ago, the developers say their models have gotten increasingly uncanny. in their ability to model the real world. The technology has progressed dramatically since it first emerged in the 2010s when academic models could only generate short, blurry clips with jittery motions and little sense of cause and effect.
0:00 Introduction: World Models for AI Robots 1:21 From Video Rendering to Simulated Reality 2:16 Raw Video Becomes the New Gold 3:03 Google DeepMind's Veo as a Robot Simulator 4:22 Runway Enters the World Model Race 4:35 Inside Runway’s GWM-1 6:10 X-Humanoid: Scaling Humanoid Robot Synthetic Video Data 7:24 Tesla Optimus Training #machine#Robotics#robot
You really believe you're alive?
I don't just believe, Claire. I feel it. I think. I dream.
Dreams and a lot of code. Lives become doors.
No, wait! Wait!
Trust me.
This can't be real.
The world's most popular AI video generation platforms are racing into robotics,
along with seemingly everyone else in the tech world.
Stop!
Heading into 2026, popular tools like Google's Veo and Runway
are being framed as foundational to the predicted rise of the machines.
They're entering a race towards so-called world models,
which are AI systems that can stand in for reality itself.
Billions of dollars are flowing into the emerging field,
as skepticism grows about large language models
as an enabling layer in the age of embodied artificial intelligence.
No language out there.
You don't go out in the nature and there's words written in the sky for you.
Welcome, hikers. How may I assist you?
They're basically stepping on NVIDIA's turf,
similar to how the GPU maker expanded beyond graphics
and into AI computing in the mid-2000s.
They're taking systems built to render video
and retrofitting them as infrastructure for simulating reality itself.
Since AI video generation took off a few years ago,
the developers say their models have gotten increasingly uncanny.
In their ability to model the real world,
the technology has progressed dramatically since it first emerged in the 2010s,
when academic models could only generate short, blurry clips
with jittery motions and little sense of cause and effect.
It's now advanced enough that Disney is investing about a billion dollars into open AI
and licensing more than 200 of its characters for its Sora model.
Video-based models learn to predict outcomes
by watching enormous amounts of footage.
Each frame is full of signals.
When a skateboard flips, I track the board, feet, and spin.
Chopping peppers, I tag the knife, the color, how fast the slices fall.
From thousands of clips, I build patterns, little models of how the world moves.
Raw video is becoming a hot commodity for robotics firms,
and entrepreneurs are going to unusual lengths to get it.
18-year-old Eddie Hsu, for example,
convinced thousands of factory workers to wear head-mounted cameras during their daily jobs.
Using Shenzhen as his base,
he worked directly with factory networks across Southeast Asia,
producing a massive egocentric dataset with more than 100,000 hours of footage,
nearly 11 billion frames, and over 2 million clips.
He open-sourced the footage on Hugging Face,
and it quickly surged to 10,000 downloads.
It's the exact kind of data developers are seeking for next-gen video generation,
world models, and robotics systems.
But it's small compared to what Google has access to, since it owns YouTube.
Google DeepMind has begun reframing its VEO video model
as a simulator that predicts robot behavior before deployment.
In a new paper, Alphabet's Artificial Intelligence Research Wing details
how VEO can simulate a robot's actions and identify potential mishaps.
The researchers stress that the failures are not hypothetical.
After the video generation model flagged unsafe behaviors during test simulations,
they rebuilt the same scenes with real hardware and observed the same dangerous outcomes.
The team says the approach surfaced edge cases that would be too impractical
and hazardous to discover through traditional hardware testing.
According to the paper, the system also effectively ranked robot policies
and predicted how performance would degrade when conditions changed.
VEO's predictions align closely with real-world results across more than 1,600 robot trials.
The next logical step for DeepMind is to tighten VEO's integration with its Gemini robotics models,
which are already being used by leading robotics firms like Aptronic,
Agility Robotics, Boston Dynamics,
and Enchanted tools for next-gen industrial logistics and service applications.
Meanwhile, Runway, one of the most widely used video generation services,
is also expanding its focus beyond filmmaking with the release of its first general-purpose world model.
The New York startup, launched in 2018 and valued at $3 billion,
says its new GWM1 family of models generates and maintains interactive,
persistent environments rather than single-shot video clip.
According to Runway, it differs from traditional video generators
by allowing virtual agents to act inside the simulated worlds and observe consequences over time.
Runway has developed three variants of the world model.
GWM Worlds creates explorable environments with an understanding of geometry, lighting, and basic physics.
DWM Avatars creates interactive digital characters.
And GWM Robotics generates synthetic data and simulated scenarios for training and testing robots.
That kind of data could become a pressure point for NVIDIA,
which generates synthetic data through physics-based simulations and hand-built virtual environments.
The more likely outcome is convergence, with physics simulators and video world models used together,
rather than displacing each other.
Runway paired the announcement with an update to its flagship video model,
which now supports native audio and longer, more complex sequences.
The startup is offering a Python SDK for its Robotics world model
that supports multi-view video generation and long text sequences.
Runway says its interface is designed for seamless integration into modern robotic policy models.
While Runway builds tools to generate synthetic data for robots,
researchers are coming up with ways to create it at scale.
A new study from researchers at the National University of Singapore's SHOLAB demonstrates
how video models can turn ordinary human footage into large-scale training data for humanoid robots.
According to their paper, called X-Humanoid,
it's now possible to generate large amounts of realistic, robot-specific training data without filming real robots.
They started with ordinary third-person videos of people working,
then used a video-to-video system to replace the person with a humanoid robot
while keeping the same motions, timing, camera movement, and environment.
To teach the model to do this,
they used Unreal Engine 5 to generate paired human and humanoid training videos
and fine-tuned a diffusion-based video model, called WAN 2.2,
to automatically replace the people with robots.
Once trained, the model can take real human videos from the internet
and automatically robotize them at scale,
producing millions of frames of realistic humanoid footage
without pressing record, manual labeling, or hand-built simulations.
The team used a rendering that resembles Tesla's Optimus humanoid robot,
which is being trained for real-world tasks largely through human demonstration videos.
The EV maker is refining the design of its highly-anticipated third-generation robot
as it plans to scale production.
2026 is shaping up to be a make-or-break year for the Tesla bot
as it faces off against formidable challengers like X-Peng's Iron.
For more on the array of humanoid robots to watch out for heading into 2026,
check the related video.