The humanoid robot industry has always had the same frustrating limitation: robots don’t just need to work — they need to be taught. And teaching them has traditionally been slow, expensive, and painfully manual. Every new task usually means custom programming, a mountain of training data, or hours of human-operated demonstrations.
That’s why 1X’s latest update for its humanoid robot NEO stands out. The company just rolled out what it calls the 1X World Model, a video-based AI system designed to help NEO turn simple voice or text instructions into real physical skills — even in situations it hasn’t encountered before. The promise is straightforward: less “engineer every step,” more “tell the robot what you want and let it work out the how.”
According to 1X, the World Model is built around video understanding grounded in real-world physics. Instead of relying only on pre-programmed behaviors or massive amounts of teleoperated robot training, NEO can learn from internet-scale video — the kind of everyday footage that shows how humans interact with objects, spaces, and routines.
The process is meant to feel simple on the user side. You give NEO a short prompt by voice or text, the robot uses its cameras to understand the scene, generates predictions about what actions should happen next, and then converts those predictions into real movement using an inverse dynamics model. It’s a big swing at one of robotics’ hardest problems: getting “digital intelligence” to translate into reliable action in the physical world.
In a recently released demo, 1X shows NEO handling everyday tasks like packing a lunch box, even when the specific objects aren’t identical to what it’s seen before. More notably, the robot is also shown doing actions it wasn’t explicitly trained on, including opening a sliding door, lifting a toilet seat, ironing clothes, and brushing a person’s hair. The company says that kind of flexibility comes from transferring broad human knowledge captured in video into robotic behavior, rather than treating each task like a brand-new engineering project.
Zoom out, and this fits a much larger trend playing out across AI right now. The frontier isn’t just chatbots producing text or models generating images anymore — it’s systems that can understand the world visually and then act inside it. Robotics has been waiting for that shift for years, because the real bottleneck hasn’t been building humanoid arms and legs. It’s been building the “brain” that can adapt when reality gets messy.
And reality is always messy. Homes have clutter, bad lighting, constantly changing layouts, and unpredictable motion from people, pets, and everyday chaos. That’s why 1X says robustness is a major focus of this update, and why the World Model is designed to keep NEO stable and responsive even when conditions shift quickly.
None of this guarantees humanoid robots are suddenly ready for mass deployment. Learning from video can help with “what to do,” but the real world still demands touch sensitivity, force control, and reliability across endless edge cases. A robot might understand what ironing looks like, for example, but doing it safely and consistently—across different fabrics, objects, and environments—is still a serious challenge.
There’s also the question of verification. As robots become more autonomous and start generating new behaviors from broad training signals, it gets harder to test every possible outcome in advance. The more flexible the system becomes, the more important safety, monitoring, and fail-safes become too.
Still, 1X’s direction is clear. The company says NEO can collect its own data as it interacts with the world, creating a feedback loop where the robot improves through experience instead of waiting for engineers to manually teach each new skill. And because the system is centered on video understanding, it could also get better as video AI models improve more broadly.
This isn’t being positioned as a far-off research project either. 1X says NEO is available through its online store, with early access priced at $20,000, priority delivery planned for 2026, and a $499/month subscription option.
For a robotics industry that’s spent years stuck in slow-motion progress, the big idea here is simple: robots have always been limited by how long it takes to teach them. If world models trained on video can actually compress that timeline and help machines generalize in real environments, humanoid robots start looking less like impressive demos—and more like products that can scale.
This analysis is based on reporting from Interesting Engineering.
Image courtesy of 1X.
This article was generated with AI assistance and reviewed for accuracy and quality.
