AGI Dinner Series | June 30, 2026
AI is getting very good at thinking on screens.
It can write code, summarize papers, operate software, and reason through complex tasks. But the next frontier is harder: getting AI to act in the physical world.
At a recent AGI House dinner, we brought together builders and researchers from Nvidia, Meta Superintelligence Labs, Recursive Superintelligence, Applied Intuition, Generalist, XDOF, Lenovo, and the AGI House community to discuss one question:
What will it take to actually deploy physical AI?
Not demos. Deployment.
Demos Are Easy. Deployment Is Hard.
The room kept coming back to the same tension: robotics demos are getting better, but real-world deployments are still rare.
In software, a bad answer is annoying. In the physical world, a bad action can break something, hurt someone, or destroy trust.
That is why physical AI cannot rely on model intelligence alone. It needs safety systems, deployment operations, monitoring, validation, and real-world feedback loops.
Self-driving cars are the warning. Waymo took nearly two decades to reach reliable deployment. The lesson is simple: the real world is not a benchmark. It is messy, high-stakes, and unforgiving.
Edge Cases Are The Product
Humans generalize from very little data. A teenager can learn to drive after a few thousand miles because they bring common sense about the world.
Today’s models still struggle with that.
A person sees police cars on the road and immediately slows down. A model may not, unless it has seen enough similar cases before. That gap matters.
For physical AI, edge cases are not rare details. They are the core challenge.
This is where reasoning, world models, and simulation may help. If a system can reason, “something unusual is happening, slow down,” it may generalize beyond its training data.
But we are not fully there yet.
Simulation Helps, But Reality Still Wins
Everyone agreed simulation is essential.
It lets robots train on dangerous or impossible scenarios: engine failures, rare road obstacles, backflips, crashes, and unusual environments.
But simulation still has limits.
Contact physics is especially hard. Cloth, napkins, soft objects, and messy household manipulation remain difficult to model. Sometimes it is cheaper and more useful to collect real-world data than to simulate everything perfectly.
The likely future is a loop:
Real robots collect data.
World models learn from it.
Simulation creates more scenarios.
Better policies get deployed.
Those robots collect even more data.
That flywheel is the prize.
The Robot Data Bottleneck
Large language models improved because the internet gave them massive training data.
Robots do not have that luxury.
High-quality robot data is scarce, expensive, and hard to collect. Teleoperation data is useful but does not scale easily. Human video is abundant but hard to transfer directly to machines. Synthetic data helps, but it still needs grounding in reality.
This is why data infrastructure may become one of the biggest businesses in physical AI.
The companies that can collect, clean, structure, and train on physical interaction data will have a serious advantage.
Will Humanoids Win?
The humanoid debate was lively.
The argument against humanoids is obvious: most great machines do not copy humans. Cars do not have legs. Planes do not flap wings. Dishwashers beat humans at one narrow task for a few hundred dollars.
Specialized robots will win many specialized jobs.
But humanoids have one major advantage: the world was built for humans.
Doors, stairs, tools, kitchens, furniture, and factories all assume human bodies. More importantly, most available training data shows humans doing things. If robots can learn from human video, then a human-shaped body may transfer better.
So the answer is probably not “humanoids or specialized robots.”
It is both.
Specialized robots will dominate narrow tasks. Humanoids are the general-purpose bet.
Physical AGI: 2030 For Capability, Later For Deployment
The group broadly separated two timelines.
Research capability may arrive this decade. Several people thought a form of “physical Turing test,” where a robot’s output becomes hard to distinguish from a human’s on certain tasks, could be possible around 2030.
Mass deployment will take longer.
Homes, factories, hospitals, and public spaces demand much more than capability. They require reliability, cost reduction, manufacturing scale, regulation, safety, and public trust.
A robot that works 80% of the time is a demo.
A robot that works 99% of the time may still not be good enough.
Safety Is Also About Trust
Physical AI safety is not only technical. It is social.
A robot may be statistically safe, but if it falls near a child, people will not care about the statistics. Public perception will shape deployment as much as engineering performance.
That means the AI community has work to do beyond building.
We need to explain what these systems can do, what they cannot do, where they are safe, and where they are not ready. Fear often comes from uncertainty. Better understanding creates room for better regulation and adoption.
The Takeaway
Physical AI is not waiting on one breakthrough.
It needs better models, better robot data, better simulation, better safety systems, better hardware, better manufacturing, and better public trust.
That is what makes the frontier so exciting.
The next generation of AI companies will not just build agents that think.
They will build systems that move, touch, assemble, inspect, deliver, repair, and operate in the real world.
The frontier is no longer just intelligence on a screen.
It is intelligence in motion.
