14 Reasons Robotics is Hard

AI is moving fast, but almost all of the visible progress so far lives inside computers. The San Francisco crowd is already looking past chatbots and toward humanoid robots that can cook, clean, carry things, and maybe even help build more of themselves. The dream is huge: AI does the thinking, robots do the physical work, and growth takes off like a rocket. But there’s a catch. Unlike ChatGPT, there’s no robot anyone can just pick up and use.

That leaves people judging progress from demo videos, which are a terrible proxy for real capability. A robot might nail one task after a hundred tries. The scene may be arranged to make the job easier. The edit may hide the pauses, the retries, the little failures that tell you what’s actually happening. Public competitions like the World Humanoid Robot Games help a bit because they’re harder to cherry-pick, but even there the source notes both impressive moments and spectacular failures.

The first big wall is the body itself. Human hands are absurdly good: roughly two dozen degrees of freedom and about 17,000 tactile sensors, all guided by a brain that can mix touch, sight, and sound. Robot hands can match one feature or another — some have as many as 27 degrees of freedom, and some fingertips pack in thousands of tactile sensors — but not the full package of flexibility, sensitivity, strength, and reliability. And the control problem is just as nasty. A robot has to grasp odd-shaped objects, fold a shirt, flip an omelette, tighten a bolt in a tight space, and react when something squishy shifts under its grip.

Then there’s perception and planning. Computer vision has improved a lot, but understanding cluttered scenes, deciding where to step, and figuring out how to reach the spice bottle at the back of the shelf is still hard. Real-world work also needs context: where supplies are kept, how a meal should be cooked, whether a nursing home resident needs help, and whether a strange gait is normal or a warning sign. On top of that, robots need to cooperate, adapt on the fly, and handle speed without turning into a liability. Today’s general-purpose robots are often slow, and the source points out that one took nearly two minutes to move to a couch, pick up one item of clothing, and put it in a washing machine.

The list keeps going: strength is hard to combine with light weight; wheels are stable but can’t do stairs; legs can climb but topple more easily; fast-moving machines are harder to stop and more dangerous when they go wrong. Battery life may be manageable, but overheating is already a serious issue, and current humanoid robots break down often. Even if the robots worked perfectly, manufacturing them at scale would take time, because the supply chains do not exist yet. So the near future of AI may be full of chatbots and agents. The physical workers are still stuck in the demo stage.