Robotics has had more false dawns than almost any technology, which makes people either dismissive or breathless. Both reactions are wrong this year, and the reason is worth understanding.
Something real changed, and it was not the hardware.
The Actual Breakthrough: VLA Models
Vision-Language-Action models are the shift. Systems such as Figure Helix, Physical Intelligence π0, NVIDIA GR00T N1 and DeepMind Gemini Robotics collapsed a problem that used to take a specialist team months.
Previously, teaching a robot a new task meant explicit programming, task-specific training and an engineer per task. A VLA model takes what it sees, an instruction in ordinary language, and produces actions. Generalisation replaced programming.
That is the same transition software went through when models started handling language instead of parsers. The hardware is not dramatically better than 2023. The brain is.
What Is Genuinely Deployed
Real deployments exist and they are narrower than the coverage suggests. BMW completed an eleven-month deployment of a Figure humanoid at its Spartanburg plant in South Carolina. Tesla Optimus, as of reporting in April 2026, is deployed within Tesla own factories rather than sold to third parties, with a stated long-term target price in the $20,000 to $30,000 range.
Notice the shape. Single-site, long-duration, heavily supervised pilots inside organisations that also build the technology or can absorb the risk. That is exactly what early industrial adoption looks like, and it is genuine progress rather than a product category.
Demo Versus Deployment
| What you see | What it usually means |
|---|---|
| A robot folding laundry | Controlled setting, curated attempt |
| A robot working a shift | Real, and worth paying attention to |
| Teleoperated demonstrations | A human is driving. Ask. |
| Multi-month plant deployment | The strongest signal available |
| A price announcement | An intention, not an invoice |
One question separates most claims: how many hours has it run unsupervised, and what was the intervention rate? Companies with good answers volunteer them.
Where the Real Money Is Right Now
Humanoids get the attention while the money sits in unglamorous specialised machines: warehouse movement, inspection, agriculture, cleaning, quality control.
These do one thing in a structured environment and they pay back on a spreadsheet. The humanoid argument is that a general-purpose body works in spaces built for humans without redesigning the facility. That argument is sound and unproven at scale.
What Is Still Hard
- Reliability over long runs. A 95% success rate is astonishing in a demo and unusable on a production line.
- Hands. Manipulation of varied objects remains the hardest unsolved problem in the field.
- Cost at volume. Target prices are targets. Manufacturing scale decides whether they arrive.
- Safety certification. The standards work is genuinely in progress, not finished.
- Integration. The robot is a fraction of the project. Fixturing, workflow and exception handling are the rest.
Conclusion
Physical AI in 2026 is real, early and narrower than the videos imply. The VLA breakthrough is genuine and it changes the trajectory, but deployments are single-site pilots with heavy supervision. If you run operations, watch this closely and pilot specialised machines that pay back today. Treat general-purpose humanoids as a two to three year planning question rather than a procurement one.
Frequently Asked Questions
Should a mid-size manufacturer be planning for humanoids?
Plan for the workflow change, not the purchase. Understanding which tasks are structured enough to automate is useful regardless of what body eventually does them.
Are humanoids better than purpose-built robots?
Almost never at the specific task. The argument for them is flexibility across many tasks and no facility redesign, which only pays off once reliability is high.
What is the single best signal of real progress?
Unsupervised runtime hours and intervention rate. Everything else is a video.