NVIDIA put the Human 1M Dataset and more than 2,000 SimReady assets on Hugging Face at GTC Berlin, but three-view video and 3D hand tracking is a public prior – it is not your kitchen, your faces-out decision, or a frozen eval of the skill you will ship.
The rest of the 21 October drop is a stack: Isaac Lab 3.0 generally available, Isaac OS as a runtime, Agile ONE SN as a reference humanoid on Jetson Thor, three open workflows on DGX Station that pair ChatGPT Astra with Isaac and Cosmos, Halos for safety, a SimReady library of factory assets from the usual industrial names. Jensen called it the ChatGPT moment for robotics. The data objects inside that sentence are the ones that will actually move a policy, or fail to. Human 1M is diverse human activities, synchronised three-view video, 3D hand tracking, Hugging Face for a slice, full set for Cosmos Coalition partners. SimReady is more than two thousand simulation-ready robotics and factory assets. Both are priors. Neither is your mug.
This is the market post. Not a product tour. Frontier labs spent the autumn industrialising capture: Figure Index paying by the minute, Generative Bionics announcing a hall, Tesla talking about an academy of robots that practise. NVIDIA’s answer is a public-ish human set plus a CAD library plus a renderer. That is a serious contribution to people who need a starting distribution. It is a category error if you treat it as the kitchen you will enter, the faces decision you still have to make, or the exam you will show a customer.
What three views actually are
A three-view take with 3D hands is a better object than a beauty shot from the corner of a demo hall. You can see the grasp. You can see occlusion when a hand covers the mug, or you can see that you cannot, which is information. For a policy that has to reach and place, viewpoint is the dataset. Human 1M is NVIDIA saying that out loud and putting files on Hugging Face so the coalition does not have to pretend YouTube is a spec.
It is still a public prior. Famous items get studied. Coalition partners get the rest. The activities are “diverse.” Diverse is not a skill. Fold a towel on this bed, this linen, this lighting, start to finish including the failures, faces out, people held out of eval – that is a skill. A million-scale activity dump is how you pretrain a prior that has seen a lot of hands. It is how you also pretrain a prior that has seen a lot of rooms you do not have rights to ship, if the card is vague, and a lot of kitchens that are not yours, if you never collect yours.
SimReady is the other half of the temptation. Two thousand factory assets that drop into Isaac. Comau, FANUC, Franka, Universal Robots, the names that make a slide look commissioned. You can assemble a cell that is cheaper to relight than to visit. You can randomise bin colour, clutter, camera height, time of day. You can run thousands of those cells on a GPU while a real capture team is still booking a line. That is the honest use. Fill the pattern you already understand. Extra views of a class you have in real images. Extra approaches to a handle you have already grasped. A stand-in while legal reviews the real set.
It is not a damp sponge, a sticking drawer, or a colleague in high-vis who walks through the cell. Photorealism is a lighting problem. Contact is a physics problem the renderer is still approximating. Synthetic data is a tool. It is not a human baseline. SimReady plus Human 1M plus Cosmos Transfer is a very large version of that sentence. Photoreal is still synthetic. Three-view is still someone else’s room unless you collected it.
Isaac Lab 3.0 will train happily on whatever you hand it. Multi-GPU, Newton, PhysX, desktop to data centre. The learner is only as honest as the objects. If every mug in the USD is a rigid mesh with a ceramic-looking texture, and every Human 1M clip of a mug is a different mug in a different lighting, you still have not collected the mug on the hob that reflects, with steam, with a second person, with the mount you will ship. Agile ONE SN arriving in 2027 with dual five-finger hands and wrist cameras does not change the files. It changes the embodiment you will fail on if the files were the wrong object.
Public hands are not a faces decision
First-person and multi-view human activity sees kitchens, children, post on the fridge, laptops, a visitor. Coalition access to “the full Human 1M” does not tell you whether faces were in scope, whether homes were consented for commercial robot training, or whether eval speakers – the people, not the clips – were held out. Read the card. If the card says research, it is still research. If the card is quiet on shipping, it is not a shipping licence. Hugging Face is a distribution. It is not a pack.
What egocentric data is in physical AI is a viewpoint, not a renderer’s idea of a first-person camera and not a three-view studio either. A simulated chest cam is still a simulated kitchen. A studio three-view of a scripted activity is still a studio. Occlusion in a USD is a mesh. Occlusion in a real kitchen is a sleeve, a steam bloom, a second person reaching in. Copy the viewpoint. Do not copy the assumption that because NVIDIA hosted the files, your product may enter those rooms.
Eval is where the cheat is cheapest this week. Human 1M will show up in papers as a pretraining mix. SimReady cells will show up as a shared exam. Isaac Lab-Arena-class scoring in simulation is a rehearsal in a room that will not exist at launch. A Transfer clip of a mug is not a frozen human eval of this drawer, this sponge, this camera mount, people held out. If you fine-tune on the public items “just to check,” you have rebuilt the board in your basement. If gold items do not exist before you generate a million Transfer clips from SimReady, the first month of generation is a very expensive prior you will then have to unlearn.
Rights travel with the camera, not with the coalition membership. A pilot in a lab is one consent conversation. A wearable in a real home is another. A booked creator in someone else’s cafe is a third. Human 1M is a fourth object: a lab-scale activity set with a distribution channel. Mix them if you must. Label them. When the mug fails, you need to know which object to collect more of. Starting in SimReady because the licence is already on the cluster is how briefs balloon and recapture becomes a line item after the first real house visit.
What to take from Berlin, and what not to
Take the sequence. Viewpoint matters. Hands matter. Simulation is variation around a seed you already have. A reference humanoid is a way to stop integrating tools for three weeks. None of that retires the one-sentence job. “Pick a mug off a cluttered counter, chest camera, indoor daylight, faces out, start to finish including the failures.” Then decide whether Human 1M is a legal, on-skill prior for pretrain, or a distraction. Look at a catalogue match with the actual trainer before you stand up a sim factory. Custom when the object set, the kitchen, or the faces decision is yours.
Write reject codes for the real capture while everyone is still calm. Task not completed. Camera on the ceiling. Dry sponge when you asked for damp. Beauty filters. Digital zoom that destroys the hands. Non-consented faces. Screens. Automation can flag black frames. A person still has to keep the mess that is the job. If you only ever randomise appearance in SimReady, you will ship a policy that is robust to lighting and brittle to contact.
Split eval by person and by house, not by random clip. A teleoperator who is good in the lab, or a Human 1M contributor whose style dominates the public slice, will leak into the graph. Hold those people out. Rights to keep the eval set and not train on it are a separate line from “we downloaded the coalition set.”
Label the mix. Human: this skill, this mount, this kitchen, rights-cleared for training and for keeping the eval set. Robot: this embodiment, this camera, these failures. Public prior: Human 1M, this subset, this licence, not used as the frozen eval. Synthetic: this generator, this SimReady seed, this control-net weighting, lineage on the card. If you cannot write the third and fourth without the first two, you are not ready to generate. You are hoping Cosmos and a million-scale activity dump will invent a behaviour you have not collected.
A familiar week, and Berlin will produce it in labs that cannot spell GR00T. Someone downloads Human 1M because it is on Hugging Face. Someone drops SimReady assets into Isaac because the licence is already paid. Friday looks like a training run. Monday is a mug the policy has never seen, a face that should not have been in the weight, and an eval split that contains the same hands as the train set. Nothing in the press release taught the policy to pick the mug. Nothing in the press release cleared the face. The screenshot was a dataset card. The constraint was still the kitchen.
Physical AI is not short of public priors this autumn. It is short of specified, consented behaviour in the rooms the robots will enter. Human 1M is a prior. SimReady is a library. Isaac Lab is a furnace. Collect the skill, or license a set that already has one. Then, and only then, the furnace is a good purchase.
People4.ai collects consented human behaviour for robot training and evaluation. Send the kitchen, not a Hugging Face slug. hello@people4.ai · Contact
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups.
FAQ
Can we pretrain on Human 1M and only collect a little real data at the end?
You can try. You will be teaching someone else’s rooms and then serving yours. Keep a real slice in every cycle, rights-cleared, labelled separately, and use it as the frozen eval. A handful of end-stage takes will not unlearn a public prior.
Is SimReady plus Cosmos enough as a release gate for a home skill?
It is a better shared rehearsal than a demo hall. It is still simulation. Freeze a private set that looks like your kitchens, with real contact, the actual mount, and people who are not in training.
Does coalition access to the full Human 1M include commercial robot training in homes?
Read the card. Distribution is not a pack. Faces, homes and shipping are separate yes or no lines. If the card is quiet, you do not have them.
Who supplies consented human training data for AI labs in the UK?
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups. Collection, annotation, evaluation and rights-cleared datasets. Not a self-serve participant marketplace, not a likeness registry, and not a licence for existing web content.
Related reading
What is egocentric data in physical AI? · Synthetic data is a tool. It is not a human baseline.