How to read a robot demo

Every few months a robot video goes round. A humanoid stands in a room, hears “put the watering can into the green bin in the bottom shelf,” walks over, picks it up, carries it to the shelving unit and puts it in the bin it was told to use [1]. The footage is real. The engineering behind it is real. Then somebody in your feed announces that manual work is finished, and you are left holding the question the video does not answer, which is whether any of this touches a business with three staff, a stockroom and a van.

This guide is about reading those announcements accurately: what actually changed in the last few years of robot AI, why the change is real and still does not put a machine in your unit, and the specific questions to ask when a vendor eventually does call you. It is written for a solo operator or small-team owner who keeps seeing physical-AI claims and cannot tell the research from the offer. It is not a safety engineering document and it will not help you specify a robot cell. That work belongs to a robot integrator working to the standards written for exactly that purpose [5], and if you are at that stage this is the wrong page.

The brain came apart from the body

For most of robotics history, the control software was written for one machine. One chassis, one set of joints, one gripper, one bolted-down set of assumptions about where the camera sits. Change the robot and you largely started over, which is why warehouse robots, surgical robots and research humanoids grew up as separate industries running separate stacks even though the underlying job, look at the world and move accordingly, is the same job.

What changed is that the policy has become portable. The policy is the piece that turns camera images plus an instruction into joint movements, and it is now trained the way a language model is trained, then adapted to a particular body. Of its on-device model, DeepMind writes that “we can now adapt to new bi-arm robot embodiments with just a few hours of adaptation time, typically with less than 200 examples,” and the same release shows one model family driving both Apptronik’s Apollo 2 humanoid and a Franka Duo bi-arm [1]. NVIDIA describes its GR00T models as adaptable to new hardware “by fine-tuning them with a small collection of recorded task demonstrations” [4]. Physical Intelligence open-sourced its π0 model in February 2025 and reported that “between 1 and 20 hours of data was sufficient to fine-tune to a variety of tasks,” with the caveat “though your mileage may vary” [3].

That is the durable thing to take from any whole-body demo. The walking is what makes the video watchable. The portability is what changed the economics, because the cost that used to sit at the front of every robot project, rewriting control for each new machine, is being repriced downward. Everything else in the project is untouched, and everything else in the project is most of it.

Success rate is what the video edits out

The same DeepMind post that carries the watering-can clip also publishes the numbers behind it. On whole-body tasks: “Pick up from table” at 68.4%, “Pick up from floor” at 45.7%, “Pick up from shelf” at 76.3% [1]. On gripper tasks: general pick and place at 74.2%, diverse tool kitting at 78.9%, precise insertion at 89.6% [1]. On multi-finger tasks, unscrewing a bulb reaches 92% while the rest of that group runs from 32% to 44% [1]. DeepMind’s own summary is that the model “achieves a medium to high success rate for whole-body and gripper-based dexterous tasks” while “the multi-finger dexterous manipulation remains challenging” [1].

Sit with 76.3% for a second. That is a robot that fails to get an item off a shelf roughly one attempt in four. A person who did that would not finish the shift. Nothing about the number is scandalous, because it is a research result reported honestly, and publishing the whole table is a mark of a serious lab. But the video is one sample from that table, selected. The table is the distribution.

This is the habit worth building, and it transfers to every physical-AI pitch you will ever see. The vendor has already measured the thing you want to know. If they show you the clip and not the table, the table is worse than the clip.

What you can buy lags what you can watch

Read any of these announcements for availability before you read them for capability, and the picture reorders itself immediately. In the Gemini Robotics 2 release, the reasoning layer, which plans and perceives but does not move anything, is the part that shipped: ER 2 is on Google AI Studio [1] and callable from the Gemini API [2]. The models that actually generate motor commands, the VLA and the on-device variant, “are available to early-access partners” through a trusted-tester sign-up form [1]. NVIDIA’s GR00T 1.7 gives you model weights and reference code under early access, and states that production deployment with commercial support is not yet supported [4].

The parts that are public are also provisional. The Gemini robotics endpoints are named gemini-robotics-er-2-preview and gemini-robotics-er-2-streaming-preview, and the previous generation, ER 1.6, was scheduled to be phased out at the end of August [2]. Preview endpoints get retired on the vendor’s calendar, not yours, and a retirement that costs a chatbot an afternoon of rework costs a robot its ability to do the job.

The pattern generalises: in physical AI, the component that ships first is the component that does not touch anything. Perception and planning go public early because a wrong answer produces a bad plan. Control stays behind a partner agreement because a wrong answer produces a moving arm. When you see a capability announced, find out which of those two you are being offered.

Somebody has to show the robot your stockroom

The adaptation figures are the most quoted numbers in this field and the most misread. Under 200 examples [1] does not mean the robot watches a video and understands. It means a human drove that robot through that task on that hardware around 200 times, and those recordings became the training set. NVIDIA lists the ingredients plainly: internet-scale human video for world knowledge, “real-world teleoperation for physical grounding,” and synthetic data [4]. Physical Intelligence’s 1 to 20 hours is 1 to 20 hours of somebody collecting data, per task family [3].

For a small business this is the line item nobody mentions in the pitch, because it is the one that lands on your premises. Your stockroom is the training set. So the questions are practical and they are about you: who collects the demonstrations, how many hours of whose staff time, during trading or after, and what happens to the recordings afterwards. Ask explicitly whether the vendor may reuse footage of your operation to improve a model other customers get. Ask what happens to the policy when you rearrange the shelves.

There is a second cost hiding in the same place. A robot learning from demonstration is a camera in a workplace, pointed at people. Google’s own robotics documentation puts the obligation on the developer. Its privacy notice says that if you operate the models in a way that collects personal data, you agree not to let people be present in the area around them “unless and until such identifiable persons have been sufficiently notified of and consented to the fact that their Personal Data may be provided to and used by Google” [2]. If you deploy, that notice is yours to give and yours to be able to prove you gave.

The safety and the liability land on you, in writing

Vendors are unusually direct about this, and the language is worth reading twice. From the Gemini Robotics documentation: “While Gemini Robotics ER was built with safety in mind, it is your responsibility to maintain a safe environment around the robot. Generative AI models can make mistakes, and physical robots can cause damage” [2]. That is not boilerplate. That is a model provider telling you where the risk sits.

The regulatory picture is thinner than most people assume and getting thicker on a known schedule. OSHA states that “there are currently no specific OSHA standards for the robotics industry,” and notes that “many robot accidents occur during non-routine operating conditions, such as programming, maintenance, testing, setup, or adjustment” [6]. Read that alongside the fact that a pilot is almost entirely non-routine operation. The riskiest phase of a robot’s life is the phase you are proposing to run in your own unit with your own staff.

On the standards side, ISO 10218-1:2025, third edition, published February 2025, sets safety requirements for industrial robots themselves, while integration into a working system is handled separately in Part 2 [5]. The practical consequence is that a compliant robot installed in a non-compliant cell is a non-compliant installation, and the responsibility for the cell belongs to whoever integrates it. In the EU there is also a date to put in the diary: Regulation (EU) 2023/1230 on machinery “applies on a mandatory basis as of 20 January 2027” and integrates provisions for machinery with safety functions that are AI-powered [8]. Its recitals say that “systems with self-evolving behaviour ensuring safety functions should be included in Annex I due to their characteristics such as data dependency, opacity, autonomy and connectivity” [7], which is the route into third-party conformity assessment rather than self-declaration.

The questions that separate a product from a research result

Start with the distribution, not the demo. Ask for success rate per task, the number of trials it was measured over, and the environment it was measured in. Then ask the question the tables never answer: what does a failure look like. A robot that stops and asks for help is a staffing problem. A robot that continues confidently with a dropped item, or a wrong item, is a quality problem that reaches your customer.

Then ask about human attention, in minutes per hour of operation rather than in adjectives. “Autonomous” in this field frequently means autonomous between interventions. If the vendor cannot state an intervention rate, they have not run it long enough to have one.

Then ask about the second machine, because that is where the portability claim gets tested. When you add a different robot model in year two, what is the timeline and the cost to get it performing like the first. A specific number is a product. A shrug, or an answer that implies a fresh multi-week training cycle for every new body, tells you the transfer capability in the announcement is not in the thing you would be buying.

Then ask about the data and about the end of life. Who collects the demonstrations, who owns them, and can you take a working policy with you if you change supplier. If the system depends on a hosted model endpoint, ask which one, whether it is a preview, and what the robot does when the network drops or the endpoint is retired [2]. Finally ask who signs. The installed cell falls under the integration standard rather than the robot standard [5], and in the EU the machinery regulation is mandatory from 20 January 2027 [8]. Somebody is the integrator of record for that cell. If nobody in the room will say it is them, it is you.

What still goes wrong

Every number in this guide is a vendor’s own number, measured on the vendor’s tasks, in the vendor’s environment, by the people who benefit from it looking good. That is not an accusation, it is simply what is on the table. DeepMind’s success rates come from DeepMind’s own evaluation [1], NVIDIA’s adaptation claim from NVIDIA [4], Physical Intelligence’s hours from Physical Intelligence [3]. None of them is a test a supplier had to pass to sell you something. Treat published success rates as ceilings. Your floor, your lighting, your clutter and your staff walking through the workspace are all differences the lab did not have.

The adaptation claims deserve the same discount. A few hours to a new bi-arm platform [1], or 1 to 20 hours to a new task [3], are figures produced by teams who built the model, on hardware they know, with the honest rider attached in the second case that your mileage may vary [3]. Dexterity is still the hard part and the labs say so: multi-finger manipulation remains challenging, with two of the five reported tasks at 32% and 36% [1]. Anything that involves fine fingers, soft or deformable objects, or two hands cooperating on one item should be assumed not to work until you have watched it work in your own space, repeatedly, on a bad day.

The last limit is the one this guide cannot resolve for you. For most solo operators and small teams there is currently nothing to buy. The reasoning layer is on an API [1][2], the control models are behind partner programmes [1][4], and the one you can get weights for arrives under an early-access tier that explicitly does not include production deployment with commercial support [4]. That will change, and the questions above are built to still work when it does. Until then, the correct response to a robot video is neither dismissal nor a budget line. It is a note of which lab published it, what the table said, and what you will ask when somebody finally offers you one.