The human body is the wrong benchmark for machines

On Saturday in Beijing, a humanoid robot called Lightning ran 100 metres in 9.32 seconds. Usain Bolt's world record, set in Berlin in 2009, is 9.58. The coverage wrote itself.

A smaller number sits further down the reports. Lightning won the Beijing half marathon in April standing 169cm tall, with legs 95cm long. Its builders at the smartphone manufacturer Honor then made the legs 10cm longer for these Games.

Bolt could not do that. Leg length is one of the few things a sprinter cannot train. Honor changed it over a summer, and the change appears on no leaderboard, because the leaderboard was built for us.

The question

When a machine breaks a human record, what has been measured?

Two answers are in circulation. Both answer the same question, and that is the problem.

The first is that this is progress, and it is. Running on two legs at 14.5 metres per second is a hard control problem. The machine is an inverted pendulum falling forward and catching itself, on hardware that heats up, inside a power budget. At the DARPA Robotics Challenge in June 2015, humanoids fell over opening doors. The rate of change since then is the finding, and the people making that argument are right.

The second is that it is a demonstration, and the conditions support that too. The Guardian's report quotes researchers saying humanoids remain mostly a matter of demonstrations, performances and research. The 9.32 came in a preparatory test event; in Saturday's final, Lightning finished second. In April's half marathon, teams followed their robots in support vehicles and had batteries hot-swapped at supply stations without the clock stopping. The standing high jump of 2.88m was reported as beating Javier Sotomayor's 2.45m, which was a running jump. The human standing best is 1.90m, set by Rune Almén in 1980. Those are two different events.

Both sides are asking how close machines are getting to us. That question is being asked loudly, and it is crowding out a better one.

What the question assumes

It assumes the human body is the destination. It is where an unguided process arrived, under constraints that have since been lifted.

François Jacob named the mechanism in Science in 1977. Evolution does not work like an engineer, he wrote. It works like a tinkerer, using whatever is at hand. An engineer can specify a part that does not exist yet and scrap the previous design. Natural selection cannot. Every intermediate form has to survive, so it cannot take a step that costs fitness now to pay later.

The consequences are visible without leaving your own anatomy. The vertebrate retina is wired backwards, the nerve fibres in front of the photoreceptors, with a blind spot where they exit. The octopus eye, which evolved separately, has neither. Our food and our air share a passage, which is why choking is possible. The nerve controlling the larynx leaves the brain, descends the neck, loops under the aorta and climbs back; in a giraffe that detour runs several metres. None of these is a design. Each is a path that could not be undone once taken.

Then the wheel. No macroscopic animal has one. Wheels are not poor: most of our civilisation rolls. But a freely rotating joint cannot carry blood vessels or nerves across it, so there is no route to a wheel along which every step is an improvement. The design was out of reach. It was never inferior. Humans built wheels five thousand years ago without difficulty.

That is what this argument has been missing. The interesting question is not how close machines are getting to us. It is what sits in the part of the design space evolution could not reach, and engineering can.

It has already started

Joints. Boston Dynamics announced its electric Atlas in April 2024 with 360-degree rotation at the hips, waist and neck, and said so in terms of not being held to a human range of motion. A machine that stands up by rotating its torso needs none of our recovery movement. It has no spine to protect.

Surgery, which left the human body behind first and with the least fuss. The da Vinci system's wristed instruments carry seven degrees of freedom inside a space narrower than a pencil, and more articulation than a human wrist. Surgeons adopted them because they do not copy the hand.

Bodies as interchangeable parts. In 2023, twenty-one institutions pooled more than a million robot trajectories across 22 different robot types and trained one policy on all of them, the Open X-Embodiment project. The shared model beat the models trained on each robot alone by roughly half again on mean success. Skill moved between machines that share no anatomy. Nothing in biology does this. What a bird learns dies with the bird.

Endurance. Lightning's half marathon was reported as a running record. The machine-native fact is the battery swap. For ten seconds the robot stopped being a closed metabolic system, then carried on. There is no human column for that, so no one recorded it.

The Fosbury Flop, and what is usually left out

The name for a change like this is the Fosbury Flop. At Mexico City in 1968, Dick Fosbury won the high jump at 2.24m going over the bar backwards while the rest of the field went face down. By Munich in 1972, 28 of the 40 competitors had switched.

Two things about that story are usually left out, and both matter.

Fosbury did not change his body. He changed the technique inside a body he was stuck with. That is the version of the breakthrough available to a human being, and it is the smaller version.

He also could not have done it a decade earlier. The flop lands you on your neck. It became survivable when universities replaced sawdust and sand pits with deep foam, and Oregon bought one. The technique was always available in principle. The environment decided when it became affordable to try.

So the question to put to robotics is not when the Fosbury moment arrives. It is: what is the foam pit?

[speculative] The likeliest answer is simulation. A robot trying an unnatural technique on a track costs money and a repair. The same robot in simulation costs compute, and can try a hundred million times overnight. The class of techniques worth attempting grows as the cost of failing at them falls. That is the same mechanism as the foam, and it has landed. Treat it as a reading of the evidence and not a settled finding: the transfer from simulation to hardware is the hard part, and it is where most of these attempts still die.

Where this was tested properly

One domain removed the human prior under controlled conditions and measured the result.

DeepMind's AlphaGo, which beat Lee Sedol in March 2016, was trained in part on a database of human games. The next version was given the rules and nothing else. AlphaGo Zero, published in Nature in October 2017, learned from self-play alone, and after three days it beat the version trained on human play by 100 games to nil. In December 2018 the same approach generalised to chess and shogi in Science.

The human data was not the foundation the system built on. It was the ceiling it had to be released from.

Games are the easy case, and the finding does not transfer on its own. They offer perfect information, free simulation and an unambiguous score, none of which a warehouse floor provides. What the result establishes is narrower and still useful. Human performance is not always the scaffolding for machine performance. Sometimes it is the constraint.

The case for building machines shaped like us

This is the strongest argument on the other side, and it is not sentimental.

The world is human-shaped. Doorways, stair risers, handles, cab controls, bench heights, the reach into the back of a shelf. Rebuilding all of it costs more than building a machine that fits. The second argument is the stronger one: the largest training set for physical work in existence is video of humans doing it, and that data is usable only by a machine with roughly human kinematics. Choose an unusual morphology and you inherit a data desert. Every example has to be generated from scratch.

That cost is real, and it is why the humanoid wave is rational. But notice what the argument is about. It concerns the legacy environment and the legacy data, not the merits of the design. It says the human form is a good adapter, and adapters are transitional by construction.

[probable] The test is what happens where the environment is new. In warehouses built for machines, nothing is humanoid: Amazon's floor drives, Ocado's grid. In surgery, the instrument beat the hand. In agriculture and infrastructure inspection, the successful shapes are the odd ones. Where the environment is old, the humanoid wins. Where it is new, it does not.

What follows

Discount the human comparisons. A record broken by a machine tells you the priorities of its designers and the rules of the event. It does not tell you about capability. Ask what the machine did that has no human column, and read that instead.

Read morphology as disclosure. When a company changes the shape of its robot, it has told you what it is optimising for. Honor added 10cm of leg before a sprint meeting. For that company, this year, the benchmark is the product.

Watch where new environments are being built. Purpose-built space, from warehouses and laboratories to fabs and clinical suites, is where non-human forms appear first. It is a better leading indicator than any games result.

The same error runs one level up, in the argument about machine intelligence. Asking whether a model reasons the way we do, or is conscious the way we are, sets the human case as the axis and measures distance along it. Those are legitimate questions and we take them seriously in The Second Look on machine consciousness. They are also, at present, close to the only questions being asked. The cost of that is a blind spot in the shape of everything a machine can do that we have no word for.

What would change our mind

If, by the end of 2028, most mobile manipulators deployed commercially in unstructured settings are bipedal humanoids, and non-humanoid designs remain confined to niches, the adapter reading is wrong. The legacy environment and the human-video data advantage would have proved to dominate, and the human form would be this generation's destination and not its bridge.

We would revise for a second reason: if transfer between morphologies stalls. The Open X-Embodiment result depends on skill moving between bodies. If each new shape needs its own data from zero, the cost of novelty stays prohibitive, and copying ourselves stays the cheap option.

Bolt's record stood for seventeen years because a human body is hard to improve. Lightning's will not stand, because a machine body is not a body. It is a draft.