But Can They Make You a Sandwich?
The Real Limits of Domestic Robots Today
Let’s start with a confession: we’ve all been impressed by videos of robots backflipping, breakdancing, or hauling crates like over-caffeinated warehouse workers. The humanoid revolution, it seems, is not just coming — it’s here! Or is it?
Despite the PR sizzle reels and billion-dollar funding rounds, domestic robots are still a long way from doing the one thing most of us actually want: being useful in our homes. Why? Because while it’s relatively easy to get a robot to move, it’s incredibly hard to get one to do something useful with its hands.
Moravec’s Paradox in Your Kitchen
Hans Moravec, one of the pioneers of robotics, pointed out something surprising decades ago: things that are hard for humans (like solving calculus) are often easy for robots, while things that are easy for us (like folding laundry or peeling an orange) are nightmares for machines. This idea, now known as Moravec’s Paradox, is more relevant than ever.
Humanoid robots can run, jump, and strike superhero poses. But can they make a peanut butter and jelly sandwich, starting with unopened jars and a sealed loaf of bread? Not without weeks of training, high-end sensors, and some human help off camera.
The Hype vs. the Hardware
Startups and tech giants alike are racing to build humanoid robots. Tesla’s Optimus. Boston Dynamics’ Atlas. 1X’s Neo. These machines look great in videos, but real-world performance still leaves much to be desired.
Take Neo, for example. In recent demos, it struggled to grip a watering can, missed a power button on the first try, and had to use its other hand just to release its grip. It can hold an egg—but not without a jittery motion that made every viewer nervous. Most of these demos rely on teleoperation or carefully edited footage. In short, we’re seeing the best attempts—not the daily reliability.
Even in factory tests—where the environment is controlled and tasks are repetitive—humanoid robots often fumble with alignment or stall out over minor changes.
What Dexterity Really Means
Robot dexterity isn’t just about moving fingers. It’s about sensing, adapting, and reacting to chaotic environments in real time.
Humans do this effortlessly. Tying shoelaces. Opening a Ziplock bag. Fishing a quarter out of a pocket. These all require flexible hands, complex tactile feedback, and fast brain loops. Most robots, by contrast, are either slow or stiff — or both.
A dexterous robot would need:
- Strong yet flexible hands
- High-fidelity tactile sensors (our hands have around 17,000 touch receptors)
- Real-time environmental feedback
- Fast, adaptive software
Even the best robot hands today (the Shadow Hand or NASA’s Robonaut 2) cost over $100,000 and can’t come close to the strength or responsiveness of a human limb. Unitree’s affordable 3-fingered hands can lift only half a kilogram.
Dexterity vs. Precision
Industrial robots have excelled at doing a few repetitive things extremely well—like placing tires on a car moving along a track. Their precision is astounding. A FANUC robot arm, for instance, has a repeatability of ±0.03 millimeters. But these machines are programmed for very specific tasks in structured environments. Change the lighting, the shape of the part, or the timing—and the robot has to be retrained.
The Hidden Complexity of “Simple” Tasks
Let’s look at just a few common household tasks that robots would likely fail at:
- Putting on a necklace with a clasp
- Baiting a fishhook with a worm
- Folding a T-shirt
- Turning to a specific page in a book
- Opening a Ziplock bag, removing one grain of rice, and resealing it
These aren’t moonshots — they’re Tuesday. And yet, they remain out of reach for even the best robots unless highly scripted and heavily supervised.
Demonstration ≠ Deployment
We’ve seen impressive robot videos before—OpenAI’s Rubik’s Cube hand, DeepMind’s origami-folding Gemini arms, or Figure’s humanoid stocking a kitchen. These demos are often breathtaking. But are they repeatable in a real kitchen with spilled coffee, variable lighting, and a curious dog underfoot?
Most of the time: no. Many demos involve either teleoperation, heavy editing, or extensive training on that specific task. Just like self-driving cars initially struggled when asked to go off-script, domestic robots have similar challenges.
The Software-Hardware Bottleneck
Dexterity is both a hardware and a software problem. Current robotic hands are weak compared to ours—NASA’s Robonaut 2 has a 20 pound grip, while the average man can lift 40 to 60 pounds in each hand. But even with better hands, robots need control systems that learn and adapt.
And that’s where things get sticky.
Modern machine learning systems depend on lots of training data. But gathering real-world tactile feedback, especially with current hardware, is painfully slow. Synthetic data can help, but it isn’t perfect. And every small change in a task (different milk carton, different lighting) can throw the whole system off.
This is why most robots are still limited to highly controlled environments. Reprogramming for variability is costly, and autonomy remains fragile. Rodney Brooks, one of the leading minds in robotics, predicts humanoid dexterity will be “pathetic” through at least 2036.
Why It Matters
If robots are going to be more than party tricks or warehouse mules, they need to get better at the boring stuff: laundry, medication, cooking even feeding the cat. The future of domestic robotics doesn’t hinge on speed or strength. It hinges on delicate pinches, smooth pours, and gentle grips.
The current trajectory suggests slow, steady improvement — not an overnight leap. And that’s okay. If your robot can carry laundry, remind you to take meds, and grab a bottle of water without smashing it, that’s still a win.
The humanoid dream might be decades off. But helpful robots that adapt, support, and make our lives easier in small, meaningful ways? That’s getting closer every day.
Robot, Android, Cyborg—What’s the Difference?
- Robot:
A machine that performs tasks, typically programmed or directed externally. Increasingly, robots are powered by AI and draw on large datasets to “learn” from experience.
- Android:
A robot designed to resemble and behave like a human, often with internal decision-making systems. Think C-3PO which even in Star Wars was limited or Data in StarTrek who was not.
- Cyborg:
A being composed of both organic and biomechatronic parts. Real-world examples include people with pacemakers or advanced prosthetics. Originally a sci-fi term (think RoboCop), it’s increasingly part of our medica
l future. The Terminator is more of an Android even though it is referred to as a Cyborg.
Coming next: Who’s really doing the work behind the scenes, how AI robots learn, and what kind of support and ethics are needed for robots to truly help us at home.
