In six days, five players in humanoid robotics have waved the same word — “autonomous”, “zero teleoperation” — as if it designated a single reality. It designates none in common. Behind the shared slogan, we find a research promise, a figure verified by a third party, a video and a closed loop. The week did not prove autonomy: it viralised the term. Here is the sorting out.
1X: “autonomous” means “learning”
On 13 January 2026, 1X Technologies unveiled the 1X World Model, a world physics model presented as the lever that would allow the humanoid robot NEO to learn new tasks autonomously. Technically, it is a generative video model with 14 billion parameters, pre-trained on large-scale videos from the internet. The stated objective is to reduce 1X’s dependence on data collection via human teleoperation: the robot would learn from unstructured video sequences, without intensive demonstrations.
This is exactly the inversion that this thread documents. For a month, the sector hid its teleoperation behind choreography; 1X, for its part, now makes it the problem to be solved publicly. But “autonomous” here means neither execution without a pilot, nor deployment: it means learning autonomy, a research trajectory. And the most spectacular claim — “NEO begins to learn on its own” — comes solely from the manufacturer and has not been corroborated by an independent evaluation. The detail that brings us back to laboratory reality: the model’s inference time is about 11 seconds per rollout at the time of the announcement. We are far from a robot that decides in real time on a production line.
Skild: 1,4 billion for a “brain” without morphology
On 14 January, Skild AI announced a Series C of 1,4 billion dollars, valuing the company at more than 14 billion dollars — more than triple its 4,5 billion seven months earlier. The product that justifies this surge, the Skild Brain, is presented as the first unified robotic foundation model capable of driving different morphologies — quadrupeds, humanoids, articulated arms, mobile manipulators — without prior knowledge of the body shape.
Here too, “autonomous” is a promise of generalisation, not a field metric. The valuation has tripled in seven months; the product remains a software model whose ability to generalise between robots is the central argument of the raise. The bar that Skild crosses is financial, not operational.
Humanoid at Siemens: the only figure the seller did not write itself
On 15 January, the contrast becomes clear. During a POC conducted at the Siemens plant in Erlangen, the robot HMND 01 Alpha by Humanoid exceeded 90% success rate on autonomous grasping and placement operations, and operated for more than 8 hours continuously. The task was not a living room demonstration: emptying bins onto a conveyor — picking up stacked bins, transporting them across the workshop, placing them for human operators.
This is the only pillar of the week where autonomy is measured by someone other than the seller: the evaluation takes place in a third-party plant, at Siemens. A rate, a duration, a real logistical task. And it is precisely this autonomy — quantified, verifiable, boring — that did not make the buzz of the week.
Galbot: “zero teleoperation” certified by the owner of the line
On 19 January, Galaxy General (銀河通用) launched the Galbot S1, an industrial dual-arm robot on an omnidirectional wheeled base. The heavy argument: the S1 would operate in full autonomy, without human teleoperation, on the production lines of CATL (宁德时代), the world’s leading battery manufacturer.
The problem is not the robot, it is the judge. The only environment where this “zero remote operation” is attested is the CATL line — and CATL is also the industrial deployer of the robot. The one hosting the demonstration is also the one benefiting from it: the sector’s “zero teleoperation” is, at this stage, self-awarded by the deployer itself. It is the closed loop that this thread has already exposed elsewhere: when the client, the host and the certifier are one and the same, the metric loses its third party.
LimX: a “world first” that is a video
On the same 19 January, LimX Dynamics published a video titled “Oli Demonstrates the World’s First Scalable Autonomous Deployment”, presented as the world first of large-scale autonomous deployment of humanoid robots. The sequence mobilises 18 Oli robots that come out of their shipping containers on their own, without any human intervention. LimX names this capability “Autonomous Awakened System” and claims it as a demonstration of large-scale multi-robot coordination intended for daily use.
It remains to be seen what the video proves. In the sequence, the 18 Oli rise autonomously, set off in coordinated formation, avoid collisions between them, then execute a synchronised choreographed routine. Yet LimX itself acknowledges: it is a capability of “manufacturing readiness” — the proof that the software architecture can manage a fleet — without any claim of real operational deployment in the field. The synchronised choreography is the show; it is not the proof. It is, word for word, the kind of “world first” that this thread has been dismantling since December.
The same word, four different weights
Put side by side, the terminological sorting is unequivocal. At 1X, “autonomous” is a learning autonomy, an uncorroborated research promise. At Skild, it is a generalisation promise backed by a valuation. At Humanoid-Siemens, it is an execution figure measured by a third party — 90%, 8 hours, a real task. At Galbot, it is a self-certification by the deployer. At LimX, it is a video that its author admits is not a deployment.
After having industrialised hidden teleoperation, the sector is industrialising its public negation. The only autonomy actually quantified by a third party does not make the buzz; the two most viral claims rest one on a video, the other on a closed loop. The question of December still holds, intact: do not ask if the robot is autonomous — ask who is holding the stopwatch.