Two announcements circulated this week on French websites. The first: one-shot imitation would now be possible for a humanoid robot. The second: surgeons would already be starting to be replaced by humanoid robots. Both rest on real, published, serious work. And in both cases, the headline asserts exactly what the paper refrains from asserting.
Nature writes 'teleoperated', the headline writes 'replaced'
On 8 July, Nature published the first in vivo feasibility study of humanoid robots in surgery, conducted jointly by the engineering school and the medical school of the University of California San Diego. Two gallbladder ablations on large non-primate mammals: the first by a human-robot team, the second by two humanoids working side by side. The machines, nicknamed Surgie, measure approximately 1.50 m for 27 kg and manipulate standard laparoscopic instruments, allowing them to fit in a cramped operating room without costly modifications. They are entirely piloted by human surgeons. The study claims no surgical autonomy, and the stated objective of the team is an autonomous assistant working alongside healthcare professionals, not their replacement. The robots had to be recalibrated several times during the procedure, and both operations took significantly longer than with an existing specialised system. The most shared French headline, however, announced that surgeons were already starting to be replaced. The surgical robot is piloted by a surgeon.
One-shot imitation exists — on arms
Same week, same mechanism. On 16 July, a team from NVIDIA, Stanford and the University of Texas at Austin submitted RoboTTT, which integrates training at test time into a robotic foundation model. The figures are spectacular and verifiable: a visuomotor context extended to 8,000 timesteps, which is about three orders of magnitude beyond the state of the art, an 87% improvement over a single-step policy, completion rates of 79%, 78% and 71% on three real tasks, and an assembly in ten steps that can last five minutes. One-shot imitation from a simple video demonstration therefore certainly exists. Simply, neither the arXiv abstract nor the project page uses the word 'humanoid': what is evaluated is a setup of bimanual manipulators. As for the French article that carried the news, it names neither the study, nor the laboratory, nor the robot, cites not a single figure and links to no publication.
What a number looks like when no one has decorated it
To measure the gap, one only needs to read, the same week, the report that Xinhua devotes to humanoids in Chinese workshops. A G2 robot from PIABOT Robotics reduces a cycle time on an automotive line from 18 to 12.9 seconds, representing about a third of efficiency gained. Aitu’s robots separate fabric pieces with 97% success, and more than 98% of the resulting sewing work passes quality control. A training centre opened in Cixi in mid-May lines up forty robots and targets more than ninety hours of effective training per day. None of these figures makes a headline. All are verifiable.
Last week, Walden Robotics raised 300 million dollars with a press release that did not write the word 'humanoid' once — the press added it. Three files, three times the same gesture: what is missing from the paper is supplied by the headline. The two distortions are not, however, of the same nature, and the distinction is worth keeping in mind. In one case, a morphology is invented: an arm becomes a man. In the other, an autonomy is invented: a piloted robot becomes a robot that decides. The latter is the most costly, because it concerns the only question that truly matters when a machine holds an instrument in an abdomen — who, exactly, is performing the gesture. The reader does not need to be sold an artificial surgeon. They need to know that a teleoperated humanoid has just removed a gallbladder, and that this is already considerable.