Before a robot touches a cup, it has to learn in a fake warehouse.

A cup slips. A cable bends. A gripper misses by a few degrees. None of those moments show up in the datasets that built the last generation of AI, and that gap is why the next one is being trained in simulation.

The text models that defined 2023 and 2024 learned from the open web. A robot cannot. The physics of holding a paper cup, dodging a human wrist, or threading a cable through a chassis does not appear in scraped Wikipedia pages or Reddit threads. To learn it, a robot either has to do it in the real world, which is slow, expensive, and sometimes destructive, or do it in software that behaves enough like the real world to teach the policy something useful. The industry is choosing the second path, and NVIDIA's overview of physical-AI simulation on Hugging Face describes how it is being built.

Real-world robot data is gathered by humans teleoperating machines through tasks, often at single-digit hertz, often in lab rooms. That pipeline cannot reach the scale of an internet scrape. A single GPU workstation running many parallel simulations can produce what the NVIDIA post calls "thousands of hours of robot experience at a fraction of the cost of real-world collection" by running the same task in many variations of a virtual warehouse at once. The robot policy that comes out of those runs has never touched a real cup, but the hope is that the physics is close enough that the policy transfers.

The pipeline that turns simulation into a deployed policy runs across three machines. The same NVIDIA post lays out a "three-computer" split: a training computer (a large GPU cluster that updates the model), a simulation computer (a GPU workstation or cluster with hardware-accelerated rendering that runs the synthetic environments), and an on-robot computer (an edge device such as Jetson AGX Thor that runs the resulting policy in the field). It is the same logic that splits a modern LLM into training, serving, and inference tiers, translated into a world where the inference target is a humanoid instead of a chatbot.

Simulation is no longer a debugging aid. Per the NVIDIA overview, it now sits inside the training loop: synthetic perception datasets, reinforcement-learning rollouts, demonstrations, data augmentation, benchmarking, and adversarial testing all happen inside the simulator. The engines doing the work are increasingly built by the same labs that train the models, and the result is that physical AI is pulling in simulation tooling the way language AI pulled in tokenizer and retrieval stacks.

Sim-to-real transfer still fails in well-documented ways. A gripper that nails a virtual cup can fumble the same cup when the lighting shifts, the foam density changes, or the human partner moves faster than the simulator expected. A GPU simulation farm isn't free; the energy and silicon cost of running thousands of synthetic hours is part of the bill. And the dataset itself is narrow by construction. A simulator that defaults to a warehouse aisle will not teach a robot how to fold laundry or catch a falling toddler, and the bias of what gets simulated becomes the bias of what gets learned.

a16z calls the distance a "physical AI deployment gap": the difference between a humanoid that walks across a stage and one that survives a 40-hour warehouse shift. Most published humanoid work is closer to the first than the second, and the gap is the strongest available counterweight to vendor optimism about how ready these systems are for real fleets.

Agility Robotics says its Digit humanoid has moved more than 100,000 totes in a working warehouse. The number is self-reported, not third-party audited, and the company is marketing it. It is also the kind of figure the deployment-gap analysis exists to measure against, and if a 100,000-tote milestone holds up under independent scrutiny, it is the first real data point on whether simulation-trained policies survive in production at all.

The same vendor sells the training silicon, the PhysX and Isaac Sim simulation engines, and the on-robot edge device, and defines what counts as a "good" simulated environment. That vertical reach across silicon, simulator, and edge is unusual, and the choice to publish the overview on a Hugging Face blog rather than a peer-reviewed venue is part of why readers should read it with the marketing caveat in mind.

The data wall that defines physical AI isn't going away. The next independent audit of a self-reported commercial number like Agility's 100,000 totes will be the first real test of whether simulation is filling the gap well enough to teach robots the things the internet cannot.