A factory robot can place a part to within the width of a hair, ten thousand times a day.
It cannot fold a towel it has never seen.
The body was solved decades ago. The missing piece is judgement in an unscripted world.
Industrial robots are one of the quiet successes of the last half century. The world average is now 162 robots for every 10,000 factory workers, more than double the figure seven years earlier. South Korea runs more than a thousand per 10,000: one robot for every ten people on the line.
Nearly all of them do the same thing. They repeat a programmed motion in a space built around them, with the parts presented in exactly the same place each time. Take away the fixture, the fence and the script, and most of them stop being useful.
The leaders run six times the world average. Most of the world has barely begun.
Source: International Federation of Robotics, World Robotics 2024 (2023 data)
A factory robot runs this loop with the world held still, so the middle steps can be written once and never revisited. A general robot has to run all of it, afresh, many times a second, in a world that keeps changing. The hardware on the outside of the loop is mature. The understanding in the middle is the frontier.
Source: Standard robotics architecture; see Siciliano and Khatib, Springer Handbook of Robotics · cited from memory, to verify
In the 1980s roboticists noticed something odd. Chess and calculus, which people find hard, were easy to program. Walking across a cluttered room and picking up a cup, which a one-year-old manages, were close to impossible. This is known as Moravec's paradox, and it still holds.
The reason is that the physical world never repeats itself. Light changes, objects slip, cloth folds a different way every time. A robot needs to perceive what is in front of it, predict what its own actions will do, and recover when the prediction is wrong. Each of those is a research field on its own.
There is also a data problem. Language models learned from the text of the internet. There is no internet of touch: no vast record of what it feels like to tighten a screw or lift a sleeping child.
Adoption is compounding at roughly 12% a year, but it is the same kind of robot, spreading.
Source: International Federation of Robotics, World Robotics reports 2017 to 2024 · cited from memory, to verify
Filled dots are tasks machines still struggle with. Almost all of them are things a small child can do. Skills that took evolution hundreds of millions of years to build feel effortless to us, and so we underestimated them. Skills we find hard, such as arithmetic and chess, are recent, and turned out to be simple to mechanise.
Source: After Hans Moravec, Mind Children (1988). Positions are illustrative, not measured · cited from memory, to verify
Three things have changed. Electric motors, batteries and sensors became cheap because phones and electric cars needed them. Simulators became fast enough to give a robot thousands of years of practice in an afternoon. And the same kind of neural network that learned language turns out to be able to learn movement, when it is shown enough examples.
The result is a new generation of machines that are trained, not programmed. They are still unreliable. A system that succeeds nine times in ten is impressive in a laboratory and useless in a kitchen. Closing that last gap, and proving it has been closed, is the work of the next decade.
This is why manipulation is the hard problem. A human hand has more independent movements and far more sense of touch than any robot hand in service, and it is run by a large share of the brain's motor cortex. Matching it takes better hardware and, even more, the ability to use it without thinking.
Source: Anatomy texts for degrees of freedom; Johansson and Vallbo (1979) for tactile receptor counts · cited from memory, to verify
Our aim is robots that can do useful physical work anywhere people do it: in homes, hospitals, fields and construction sites, and eventually in places people cannot go. We care most about the work that is dangerous, exhausting or simply not being done because there is nobody to do it.
We intend to work in the open. That means shared datasets of real manipulation, reference designs that a university lab or a small company can build, and honest benchmarks: how often does it fail, and how badly. A robot that shares a room with a person must be safe by construction, and its limits must be published.
Move your pointer over the stage. You only say where the hand should go. The arm works out the three joint angles itself, many times a second. Then let it work alone.
This is a statement of intent, not a record of achievement. Each step carries a label for how firm it is. Nothing marked proposed or speculative has been done yet.
A fourth platform beside Anu, Dhi and Synapse: actuators, sensing, control and learning, each as a working simulation.
Recordings of real tasks with force and touch, plus a public test of reliability that anyone can run.
A documented, affordable arm and hand for research groups that cannot buy their way in.
A machine that completes a real, unscripted task (for example, restocking a ward) with a published failure rate.
Machines that learn a new physical task from a demonstration and work safely beside people, on Earth and off it.
Every claim on this page is open to challenge in the Agora, in the Computing & Hardware room.
Argue it in the Agora ↗