There’s a palpable excitement in Silicon Valley right now. Tesla is building humanoid robots. Google DeepMind launched Gemini Robotics. Startups are raising millions on pitch decks about robots folding laundry. The word “embodied AI” is suddenly everywhere.

When I started working in robotics again, after years of watching the field promise breakthroughs that never quite arrived, I was skeptical. I’d heard “this time is different” before.

So if you’re outside this world, you might be wondering: what’s the big deal? We’ve had robots for decades. Factories have used robotic arms since the 1960s. Amazon warehouses are full of autonomous machines. So why does 2026 feel different?

The answer lies in understanding what robots couldn't do until now—and what's suddenly changed to make the previously impossible feel within reach.

The Long History of Robots That Could Only Do One Thing

To understand where we are, we need to understand where robotics has been.

The first industrial robots were essentially programmable machines that followed fixed paths. You could program a robotic arm to weld a car door in exactly the same spot, thousands of times a day, with incredible precision. But here's the catch: the environment had to be perfectly controlled. The car door had to arrive at exactly the right position, at exactly the right angle, every single time. The robots were caged off from humans. Any deviation from the script, and the whole thing fell apart.

This worked beautifully in manufacturing. When you're making a million identical parts, you can afford to engineer an environment where nothing is ever a surprise.

Then came the next wave: robots with computer vision and control systems that could handle some uncertainty. Amazon's warehouse robots are a good example. These machines can perceive their surroundings, avoid obstacles, plan a path across a warehouse floor, and respond to changes– like a human walking past. This was a genuine breakthrough. The robot didn't need a perfectly controlled environment anymore. It could adapt, at least a little.

This approach is still being pursued today, and for good reason. There are plenty of applications where a robot needs to do one thing, or a handful of things, reliably for eight hours a day. In those cases, the classical approach works well.

But here's the problem: the real world is 99% not like that. Consider a McDonald’s associate. They aren't programmed for a single, repetitive loop. In the span of ten minutes, they might drop a basket of fries, bag an order, mix a shake, and then jump on a register to help a confused customer. They are constantly switching contexts. Furthermore, every one of those tasks is full of dynamic, unscripted variables: a customer changing their mind, a spilled drink, a fryer that’s running too hot. You can't plan a rigid path for that kind of reality.

What Humans Can Do That Robots Couldn't

Think about what you do with your hands on any given day. You tie your shoes. You fold a shirt. You pick up a single sheet of paper from a stack without grabbing the whole pile. You thread a needle. You crack an egg. You adjust your grip on a slippery glass.

None of these tasks require superhuman precision. In fact, human limbs aren’t particularly accurate or repeatable-- some robot arms can place objects with micron-level precision, something an ordinary person could never do without years of specialized training.

But humans can do something robots have never been able to do: we can handle novelty. We can walk into a kitchen we’ve never seen before and make breakfast. We can figure out how to open a jar we’ve never encountered. We can adapt on the fly when things don’t go as expected.

This ability-- call it common sense, or generalization, or adaptive intelligence—has been the Holy Grail of robotics for decades. It’s the reason you still don’t have a robot that can clean your house the way a person can, even though we’ve had Roombas for twenty years.

The question that’s kept researchers awake for generations is: how do you give a machine the kind of flexible, adaptive intelligence that lets a five-year-old figure out how to use a new toy?

What ChatGPT Has to Do With Robots

Here’s where, even as a skeptic, the story gets interesting.

If you’ve used ChatGPT, you’ve noticed something remarkable: it sounds human. Not perfectly human, not always, but close enough that you sometimes forget you’re talking to a machine. It can write poetry, explain quantum physics, help you debug code, or plan a dinner party. It seems to have something like common sense.

How does it work? At a fundamental level, ChatGPT is trained on an enormous amount of text from the internet– billions of pages of books, articles, conversations, code, and everything else humans have written. The AI learns patterns. Given a sequence of words, it learns to predict what word might come next. Do this at a massive scale, with enough data and enough computing power, and something surprising emerges: the model starts to behave as if it understands language. It can generalize. It can handle situations it’s never explicitly seen before.

This is the breakthrough that’s changing robotics.

Because it turns out that what works for text also works for other kinds of data. The same fundamental approach has been extended to images, sounds, videos, and three-dimensional environments. If you can convert something into numbers– into data that a neural network can learn from-- you can apply the same techniques.

And robots? Robots are just another kind of data.

Training Robots Like We Train Language Models

Here’s how it works.

A robot has motors that control its movement. It has cameras (or other sensors) that let it see the world. Both of these, motor positions and camera images– can be converted into numbers. The robot’s experience of the world becomes data.

Now, imagine you have millions of examples of a robot performing a task. Each example includes what the robot saw (the camera images) and what the robot did (the motor commands). Train a neural network on this data, and it learns the relationship. Given a particular scene, it learns what actions lead to success.

This is exactly what Tesla is doing with its self-driving system. They’ve replaced their entire software stack: perception, planning, execution, with a single large neural network. That network was trained on data from their fleet: millions of miles of driving, billions of camera images paired with human driver inputs (steering, acceleration, braking).

The AI learns to mimic human drivers. And with techniques like reinforcement learning, essentially a system of rewards and penalties that teaches the AI which behaviors to copy and which to avoid, you can make it mimic good drivers specifically.

Tesla’s approach works because driving, as complex as it is, has a useful property: the car’s actions don’t fundamentally change the environment. You steer, you accelerate, you brake, but the road stays the road.

In robotics, this is known as locomotion. It's simply about moving the machine from Point A to Point B, where the world is essentially a backdrop to be navigated.

Manipulation is different. It is about reaching out and changing the world. It’s grasping, twisting, folding, and rearranging objects. It is the difference between walking past a messy room and actually cleaning it up.

The Hard Problem: When Robots Change the World

When a robot picks up an onion and slices it, the world changes. The onion goes from whole to diced. When a robot folds a towel, the towel goes from flat to folded. Every action the robot takes transforms its environment in ways that affect what it needs to do next.

This creates a data problem. To train a robot the way we train a language model, you need enormous amounts of demonstration data – recordings of robots (or humans) performing tasks, captured in a way that includes both what they saw and what they did. For language, this data existed. The internet is a nearly infinite source of text. For driving, Tesla could collect data passively from their existing fleet.

For manipulation? We're starting from almost nothing.

This is where the field is focused right now. Researchers are attacking the data problem from multiple angles:

• Real-world data collection: Building systems to efficiently capture human demonstrations of physical tasks. This means figuring out how to record not just what a person's hands are doing, but how to translate that into commands a robot can execute.
• Simulation: Creating virtual environments where robots can practice tasks millions of times without wearing out physical hardware. The challenge is making simulations realistic enough that skills transfer to the real world.
• Learning from existing video: There are billions of hours of humans doing physical tasks on YouTube. Can we extract useful training data from videos that were never intended for robot learning? This is an active area of research.
• More efficient learning: Making AI systems that can learn from fewer examples. Humans don't need to see a thousand demonstrations of how to pick up a cup; we figure it out from a handful of tries. Can robots do the same?

Why Hardware Is Getting Cheaper

There's another part of this story that doesn't get enough attention: the hardware is changing too.

Traditional industrial robots are precision machines. They're expensive because they need to be incredibly accurate and repeatable. When your robot arm needs to place a component within microns, you need high-quality motors, precision gearing, and sophisticated control systems.

But if your AI can handle imprecision, if the intelligence is smart enough to compensate for hardware that's "good enough" rather than perfect, suddenly your hardware requirements drop dramatically.

This is the bet a lot of companies are making. The AI handles the tolerance and adaptability; the hardware just needs to be adequate. You don't need a $100,000 industrial robot arm when a $10,000 arm paired with smart enough software can accomplish the same task.

This coupling: sophisticated AI with less precise but much cheaper hardware, is what could make robots economically viable for applications that were never possible before.

What This Could Mean for the World

Let me be clear: we're not there yet. The challenges are real, and the timeline is uncertain. But the excitement in the field is genuine because, for the first time, the path forward is visible.

If robots can develop something like common sense—the ability to handle novel situations, to adapt on the fly, to work in environments that haven't been perfectly engineered—then suddenly the applications expand enormously.

• Senior care: Right now, caring for aging populations is labor-intensive and expensive. Robots that can assist with daily tasks could make quality care more accessible.
• Food service: Restaurants struggle with labor costs and consistency. Robots that can handle the variability of a real kitchen—not a perfectly controlled production line, but an actual working kitchen—could transform the economics of food preparation.
• Disaster response: When buildings collapse or floods hit, first responders face dangerous and unpredictable environments. Robots that can navigate rubble, adapt to damaged infrastructure, and perform physical tasks in chaotic conditions could save lives.
• Construction: Housing costs are driven in part by labor costs. Robots that can perform building tasks—not just in factories making prefab components, but on actual construction sites—could make housing more affordable.

The thread connecting all of these is the same: they require intelligence that can handle the real world in all its messy, unpredictable complexity. That's what's been missing. That's what might finally be within reach.

Why This Matters

There’s a phrase that gets thrown around in robotics circles: “physical intelligence.” It captures something important. The breakthroughs in AI over the past few years have been about intelligence that operates in the realm of language and images and code-- things that exist as information. Physical intelligence is about extending that to the real world. It’s about AI that can not just think, but do.

If this works-- if the approaches that made ChatGPT possible can truly be extended to robots operating in the physical world-- we’re looking at one of the most significant technological transitions in human history. Not because robots are new, but because robots that can adapt, generalize, and handle novelty would be new.

People talk about a future of sustainable abundance. It’s a vision where the basic necessities of life-- shelter, food, care-- become more accessible because the labor required to provide them can be augmented or replaced by machines. Whether or not you believe we’ll get there, and how long it might take, the pursuit is underway.