Physical AI: Where It Actually Stands, and Where It's Going

AI learned to write essays before it learned to fold a towel.

That's not an accident. Text is cheap, abundant, and forgiving: a wrong word costs nothing. The physical world is the opposite. Data is expensive to collect, every action has consequences, and gravity doesn't accept a retry. A language model that's 95% right is useful. A robot that drops 5% of what it picks up is scrap.

Physical AI is the effort to take the learning methods that cracked language and vision and point them at bodies: arms, legs, wheels, rotors. It's the most interesting frontier in AI right now, and also the one with the widest gap between demo videos and deployed reality.

What Physical AI Actually Means

Classical robotics is hand-written. An engineer models the arm, plans a trajectory, tunes a controller, and the robot does that one thing very precisely, forever. Change the part, the lighting or the bin, and someone rewrites the code.

Physical AI flips that. Instead of programming behaviour, you learn a policy: a model that takes in camera frames, joint states and often a language instruction, and outputs motor commands directly. The same model that stacks blocks can, in principle, be told to clear a table.

The interesting part isn't any single task. It's the promise of generality: one model, many bodies, many jobs.

move your cursor ↙

θ1 0.0°θ2 0.0°x 0y 0

Where It Stands Today

The honest picture is split in two.

What works in production:

• Autonomous driving, in bounded areas. Waymo runs fully driverless robotaxis in several US cities. It's the clearest proof that learned perception and planning can operate in the open world — and it took over fifteen years, enormous fleets of data and carefully mapped service areas to get there.
• Warehouse and logistics automation. Mobile robots moving shelves and totes are a mature industry. Learned grasping now handles the messy part, picking mixed items out of bins, at real commercial volume.
• Drones. Learned perception and control run inspection, mapping and agricultural spraying every day. The physics is simpler than manipulation — mostly avoid touching things — which is exactly why drones got there first.

What's working in pilots, not at scale:

• Humanoids. Figure, Agility, Tesla, Apptronik, 1X, Unitree and Boston Dynamics all have humanoids doing real tasks in real facilities: moving totes, loading parts, sorting packages. But "deployed" usually means a handful of units, on narrow tasks, with people close by. Nobody is running a factory on humanoids yet.
• General-purpose manipulation. Robot foundation models — Physical Intelligence's π models, Google DeepMind's Gemini Robotics, NVIDIA's GR00T — can fold laundry, bus tables and follow instructions they weren't explicitly trained on. That's a genuine leap from five years ago. They're also slower than people, and still fail in ways that are hard to predict.

The pattern is the same one I see in software agents: the capability is real, the reliability is not there yet.

Why It's So Much Harder Than Language

Four problems make physical AI a different discipline, not just a harder benchmark.

1. There's no internet of robot data. Language models trained on trillions of tokens that already existed. Nobody has trillions of examples of hands opening drawers. Every demonstration has to be collected with teleoperation, on real hardware, by a person. The largest robot datasets are still tiny next to text corpora.

2. Simulation lies a little. Simulators can generate unlimited data, but friction, contact, cable sag and deformable objects never behave exactly like the real thing. Policies that look perfect in sim fall over on hardware. Closing this sim-to-real gap is half the field's work.

3. Mistakes are physical. A hallucinated sentence is embarrassing. A hallucinated grasp breaks a glass, a gearbox or a finger. Safety can't be a filter on the output; it has to sit in the control loop, with hard limits the learned model can't override.

4. Hardware is slow and expensive. You can't spin up ten thousand robots overnight the way you spin up GPUs. Actuators wear, batteries limit shifts, and every hardware revision invalidates some of your data. Software iterates in hours. Hardware iterates in quarters.

I learned a small version of this building a robotic arm and a drone. The model was never the hard part. Calibration, latency, backlash and the one sensor that drifts when it warms up — that's where the weeks went.

What's Changing Right Now

Three shifts are pushing the field forward faster than at any point I can remember.

Vision-language-action models. Instead of training a policy from scratch, you start from a model that already understands images and language, then teach it to output actions. The robot inherits common sense for free: it knows what a mug is, that cups go upright, that "the red one" means something. This is the single biggest reason recent demos look so different.

World models. Models that predict what happens next in a scene — NVIDIA's Cosmos and DeepMind's Genie line are examples — can generate training data and let robots rehearse before acting. If they get accurate enough, they turn the data problem into a compute problem, and compute problems get solved.

Cheaper hardware. Capable humanoids and arms have dropped sharply in price, driven largely by Chinese manufacturers. Cheap hardware means more robots, more robots mean more data, and more data is the bottleneck. That flywheel has started turning.

Where It's Going

My honest read on the next few years, from nearest to furthest:

Next 1–2 years. Humanoids and mobile manipulators move from pilots to small fleets in structured settings: warehouses, automotive plants, logistics hubs. Tasks stay narrow, but they stop being staged. Robotaxis expand to many more cities. Foundation models become the default starting point for any new robot project, the way pretrained language models did for NLP.

3–5 years. Robots that can learn a new task from a few demonstrations instead of thousands. Real deployments in retail backrooms, hospitals and commercial kitchens — places that are semi-structured but full of variation. The winning companies won't have the best demo; they'll have the largest fleet feeding data back into training.

Further out. Home robots that are genuinely useful, not just novel. This is the hardest environment there is: every home is different, full of soft and fragile things, and the tolerance for failure is near zero. I think it happens, but later than the hype suggests, and it arrives first as narrow helpers — laundry, tidying, elder-care assistance — rather than a do-everything butler.

The Real Bottleneck

The headline race is about models. The real race is about data and deployment.

Language AI scaled because the data already existed. Physical AI only scales if robots get deployed doing useful work, collect data while they do it, and feed it back into better models. That loop — deploy, collect, retrain, redeploy — is the whole game. A company with a slightly worse model and ten times the fleet will win.

That also means the unglamorous engineering matters more than ever: reliable hardware, safety systems that hold under edge cases, teleoperation tools that turn a human fix into a training example, and observability for every action a robot takes.

What Comes Next

We're roughly where language models were just before they became products: the capability is clearly real, the reliability isn't, and the people closest to the work disagree about the timeline.

What I'm confident about is the direction. AI is leaving the screen. The next decade of the field won't be measured in benchmark scores — it'll be measured in boxes moved, meals made, crops sprayed and hours of human work given back.

The digital revolution changed how we think. The physical one will change what gets done.