The week of 10 to 16 March 2026 will be remembered as the week of the humanoids: a giant training hall in Munich, an NVIDIA keynote at GTC, a Chinese five-year plan, a puppet-robot on stage, a projection of three billion machines by 2060. Six announcements, six headline acts, and the temptation to file them into camps — Europe versus China, open source versus proprietary. That is to miss the one sentence that ties them together, and it was NEURA Robotics that wrote it in black and white: the main bottleneck in robotics is no longer the hardware, it is the acquisition of large quantities of real physical training data.

Once that sentence is laid down, the week stops being a juxtaposition. It becomes a battle — with three incompatible answers — over a single question: where does the data that trains robots come from? Gather it by hand, synthesise it by computation, or decree it from the state. Nobody won. The fight was over who owns the factory.

Answer no. 1: gather it by hand

On 10 March, NEURA Robotics and the MIRMI of the Technische Universität München jointly announced the TUM RoboGym. The pitch is blunt: the largest scientific centre in Europe dedicated to training Physical AI at the time of its announcement. But the set-up inside says more than the superlatives about what is really going on. Training data is captured there by human trainers wearing MANUS gloves and an Xsens full-body motion capture system, then fed into the Neuraverse platform.

This is the literal admission of the bottleneck: humans are paid to make, by hand, the raw material that is missing. Neuraverse is in fact conceived as a hardware-agnostic system for collection and distribution, meant to open up to industrial partners and startups, and the majority of the data generated at the RoboGym will be shared as open source with the robotics community. The human gesture, captured and redistributed, as a resource. Data is not a by-product of training: it is the product, and it takes a 17-million-euro hall to produce it.

Answer no. 2: synthesise it by computation

At GTC 2026, on 16 March, NVIDIA does exactly the reverse — and owns it. Jensen Huang gives a first look at GR00T N2, a robotics foundation model born of the internal DreamZero research, with a "world action model" architecture, whose general availability is announced for the end of 2026. According to NVIDIA, the robots running it succeed at new tasks in unfamiliar environments more than twice as often as rival vision-language-action models.

Around the model, a whole synthesis chain: the Physical AI Data Factory Blueprint, an open reference architecture for generating training data at scale, built on the Cosmos world model and the OSMO orchestrator; Isaac Lab 3.0 in early access, a reinforcement learning engine resting on the Newton 1.0 physics engine and optimised for DGX infrastructure. The Decoder sums up the manoeuvre without mincing words: NVIDIA wants to "turn robotics' data problem into a compute problem".

The demonstration that condenses it all is Olaf. To train this robot, 100,000 virtual instances were simulated simultaneously in two days on a single NVIDIA RTX 4090 GPU, via the Kamino simulator — a proprietary tool from Disney Research, built on the Newton solver and GPU-accelerated. There lies the hard figure, dug out from under the folklore of three billion humanoids: not a 2060 horizon, but 100,000 instances in two days on a consumer card. The day training a robot stops being a problem of arms and gloves and becomes a problem of graphics cards, the industry's centre of gravity leaves the workshop for the data centre. And the consequence is plain: if data becomes compute, whoever sells the compute owns the data by default.

Answer no. 3: decree it

On 12 March, the National People's Congress formally adopts China's 15th Five-Year Plan (2026-2030). For the first time, embodied AI is classed there as a strategic industrial category in its own right, on a par with quantum, brain-machine interfaces and 6G; humanoid robots feature among the eight designated national strategic industries, a status upgrade from the 14th plan.

The detail that matters here is not the general ambition, but the list. The plan sets out five axes for embodied AI, and two of them name data precisely as state infrastructure: the coordination of training grounds and the virtual-real fusion for collaborative training — that is, the first two answers, gathering and synthesis, planned from the top down. All of it backed by a 1,000-billion-yuan state venture capital fund (about 138 billion dollars) dedicated to AI, robotics and emerging technologies.

And for the data to exist, you need bodies that produce it. In March, UBTech Robotics signs a strategic cooperation agreement with Siemens Digital Industries Software to scale production of the Walker S2 to 10,000 units a year, integrating Siemens' design, simulation and planning software across the entire manufacturing cycle. Mass-producing the machines that, once deployed, will in turn generate the data: the loop closes at the scale of a state.

The "geopolitical bifurcation" doesn't hold

The easy narrative of the week was one of a rift: a European counterweight — the RoboGym explicitly positions itself, moreover, against the forty Chinese collection centres and the American initiatives — against rival blocs. The week's graph dismantles that narrative.

The only real autonomy demonstrated, with no teleoperator, was at GTC: RealSense and LimX Dynamics carry out the world's first demonstration of autonomous 3D navigation for a humanoid, combining RealSense depth cameras and NVIDIA's cuVSLAM, the robot moving through spaces shared with humans without the slightest operator intervention. Yet this robot is Chinese — LimX is based in Shenzhen — and it was trained by reinforcement learning in NVIDIA Isaac Lab to close the sim-to-real gap.

That same week, NEURA (Europe), Figure AI (United States), Agility Robotics (United States) and AgiBot (China) all announce the adoption of the Isaac GR00T N models — while NVIDIA hands the Newton engine to the Linux Foundation and releases it as open source under the Apache 2.0 licence, co-developed with Disney Research and Google DeepMind. The "camps" may be fighting over the data by geography. But they all train it on the same simulator, with the same foundation model, on the same physics engine. The rift is in the collection, not in the tool: at the level of the stack, there are not two blocs, there is one supplier.

One floor lower

This thread descends the stack layer by layer, week after week. In December, the demo was the product; at January's CES, teleoperation was becoming a data business; in February, the body turned out to be consumable, the value lodged in the brain; on 9 March, the question was: who owns the brain? This week goes one notch further down, beneath the brain: who owns the data factory that feeds it? The object is new — the training data and its mode of production — and it is attested in the same week by a physical centre (the RoboGym), a synthesis stack (the Data Factory and Olaf) and a state plan (the 15th Plan).

So don't ask which robot performed best this week. Ask where its training data comes from. The three answers — gather it by hand, simulate it, decree it — draw the real map of forces. And while the three are busy at it, NVIDIA is working to make sure only one question remains to be asked: how many GPUs?