Why Robots Still Can't Do the Simple Stuff (and What's Changing)

Moving is solved. Touching isn't.

We've gotten really good at making robots move. Legged robots, wheeled platforms, and arms with millimeter repeatability all exist and work. But a robot that can get anywhere and can't pick anything up is basically a very expensive tourist.

That's why so much of the real value is in dexterity: gripping, holding, inserting, adjusting. It's also why many teams now reuse proven arms and bases instead of designing a whole new robot. If the hardware already works, don't reinvent it. Put the effort into the hand and the intelligence behind it.

Learning from the web only gets you so far

A common approach to robot AI is to take models trained on internet text and video, then fine-tune them for motor control. On paper it makes sense, since those models know a lot about the world.

But reading about skiing doesn't make you a skier. Text and video can't tell a robot how slippery a metal surface feels, how a cable resists bending, or when a smooth plastic part is about to slip out of its grip. Those things only show up in real physical interaction data, so more of the field is moving toward training on that directly. Turns out the internet is great at cat videos and terrible at friction.

Collecting that data is the real challenge

Teleoperation, where a human drives a robot with controllers, is accurate but slow, expensive, and stuck in the lab. The alternative is wearable devices like cheap gloves that let people do normal tasks while recording hand motion and contact.

That means you can collect huge amounts of data in homes, workplaces, and all kinds of environments instead of one spotless lab. Diversity matters here. A model trained only on tidy lab conditions will panic the first time it meets a messy desk.

From rigid scripts to improvisation

Traditional industrial automation follows a fixed trajectory. If everything is exactly where it's supposed to be, it works great. If a part shifts a few millimeters, the controller leaves its expected path and gives up, like a GPS that never says "recalculating."

Newer learned systems can actually recover. Bump the robot, shift the part, change the lighting, and a good model adjusts mid-motion and keeps going. Sometimes it even does things it was never explicitly shown, like switching to the opposite hand. This is where robotics starts to feel less like programming and more like teaching, except the student never complains about homework.

It also changes how you fix problems. Instead of rewriting control code for a tricky task like inserting a flexible hose, you demonstrate the right motion, retrain, and let the model absorb it. Some people call this "programming through data." I call it the first time "just show it how" is a legitimate engineering strategy.

Demos aren't products

Robotics has a demo problem. A polished video of a robot folding laundry under perfect lighting tells you very little about whether it works in a factory. The better test is whether it can handle real parts and real tasks that a business would pay to automate, including when conditions change and nobody is filming.

Real deployments also come with unglamorous constraints:

• Bad connectivity. Warehouses often have weak or no network, so you can't count on fast cloud uploads. Your robot's brain can't depend on Wi-Fi that your phone can't find either.
• Customers want a direct benefit. Companies rarely care about improving someone else's AI model. They care whether sharing data makes their own workflow better.
• Hardware variety. Models that work across different robot arms and grippers are far more useful than ones locked to a single machine.

My takeaway

The future of robotics probably isn't about more impressive hardware. It's about giving robots a feel for physics, learned from real interaction, so they can handle the messy, unpredictable situations that scripted systems can't. Mobility got robots to the job site. Dexterity and adaptability are what will let them do the work once they're there, and maybe even fold my laundry.