Learning by falling

Most people will meet humanoid robots first as spectators and later as customers. What they learn in between will matter more than any record set on the track

A humanoid robot ran 100 meters in 8.64 seconds at the Second World Humanoid Robot Games in Beijing last month. A year ago, at the first edition, the same event was won in 21.50 seconds. The numbers went around the world in hours.

I was in the stands for much of it, and they are not what I have been thinking about since. What stayed with me are three shifts that no stopwatch can record: in what an ordinary crowd will feel about machines, in what those of us who build them should ask of them, and in where the field's knowledge has gone to live.

Start with the stands. Between records, robots tripped, collided, seized up and were carried off on stretchers; one caught fire. The crowd laughed, then cheered.

Few people in that stadium build robots for a living, and that is why the format matters. Sport is an old, efficient technology for making capability legible to non-experts. You need no paper to see who crossed the line first, or that a machine which can sprint cannot yet climb a spiral staircase unaided.

What the audience mostly judged was failure, and that is the valuable part. Trust in machines is not built by demonstrations of competence; it is built by exposure to incompetence. A technology met only through launch videos gets worshipped or feared, because without evidence people fall back on hope or dread. A machine you have watched trip over its own feet becomes something better: an object you can estimate, price, regulate and eventually buy.

The clip that traveled furthest was no record. A small robot won its 400-meter final while appearing to cover its face with both hands, and the internet decided it was shy. Asked how they taught it that, its developers said they had not: the posture emerged from optimization, an arm motion that helped it keep balance.

The system was solving for stability. The audience saw a personality. That gap, between what a machine is doing and what people believe it is doing, will shape this technology's politics and commerce more than any capability curve. It cannot be legislated away, because reading intent into moving things is a default human behavior, not ignorance. Most people will meet a humanoid first as spectators and, within a decade, as customers. The model they carry between those moments is not a soft question. When a chatbot is wrong you get a bad sentence. When something with mass and momentum is wrong, it falls on something.

That brings me to the reading I would most like colleagues abroad to revise. Coverage split in two: one half took the clock as proof that a race had been won; the other took the stretchers as proof that none of it was serious. Both are wrong in the same way: They mistake a scoreboard for the exam.

An obvious question about a humanoid is why we insist on the human shape. The answer is unglamorous, and it also answers the replacement anxiety. The world is designed for human dimensions: stair risers, door handles, tool grips and bench heights. Retrofitting the machine is cheaper than retrofitting the planet. The humanoid form is backward compatibility with a world we have finished building; it is about accommodation rather than about manufacturing people.

The competition understood this better than most commentary. Take the same 100 meters, add 10 obstacles, and the winning time was not 8.64 seconds but 1 minute 28.32, in one of the few track events still permitting remote operation. In the office scenario, robots had 20 minutes to load a printer, set out meeting supplies and run a shredder, with full autonomy scoring highest but human involvement not disqualifying. The rules had encoded the real question. It is not whether a person remains in the loop. It is where.

That is the engineering problem of the coming decade, and a better one than autonomy. A machine that fails safely and legibly, that knows when to stop and ask, is worth more in a hospital corridor than one that succeeds three percent more often. For a decade we asked what robots can do. The better question is what we should keep. Constancy and patience for the dull and the dangerous can be handed over without regret; intent, judgment and responsibility for the exception cannot. That is no country's problem to settle alone.

The third shift I noticed lies in my own profession. Young researchers now say something unsayable when I was their age. Doctoral students, postdoctoral fellows and undergraduates at Peking University's Tong Class (Gifted Program in General Artificial Intelligence) offer versions of it: A week on a competition floor teaches them more about the state of the art than a week at the International Conference on Robotics and Automation or the International Conference on Intelligent Robots and Systems, the field's flagship international conferences.

My instinct was to correct them. I was formed by that system and still publish in it. But they describe something real, and it is not about prestige. A paper is a lossy compression of an experiment. In embodied intelligence, nearly everything that matters lives in what it discards: How a foot slips on one floor surface, why the software that keeps a robot upright works for one hour but sometimes not for three seconds. Such knowledge does not travel in PDFs. It travels in people, in parts and in repeated contact with the physical world, so the field's center of gravity follows that contact, wherever it is densest. That is a structural claim, not a national one: it has been elsewhere before and could be again.

Over 2,000 machines from 16 countries came to Beijing to be measured against each other. The flat 100 meters has been won many times over. The one with obstacles has been won by no one. That race needs no further record, only shared yardsticks: common benchmarks, incident reporting and a common definition of what counts as autonomous. A record can only be admired from a distance. A test can be run by anyone, anywhere, on their own machines. That is the invitation worth extending.

The author is an assistant professor at the School of Psychological and Cognitive Sciences and the Institute for Artificial Intelligence at Peking University.

The author contributed this article to China Watch, a think tank powered by China Daily. The views do not necessarily reflect those of China Daily.

Contact the editor at editor@chinawatch.cn.