Point of view

Film, video and 3D for physical AI companies

Physical AI is the deep tech category whose claim lives in motion, which is why a photograph and a spec sheet cannot carry it and a film usually has to.

Every deep tech category has a legibility problem. Physical AI has a different one: the product is not a thing, it is a sequence, and almost nothing a company publishes is capable of holding a sequence.

Why can a photograph not sell an autonomous machine?

Because the photograph shows the housing, and nobody is buying the housing.

A still of a robot arm above a conveyor tells a buyer the arm exists, that it is a certain size, and that somebody has paid for good lighting. It says nothing about the thing being sold, which is what happens in the four seconds after a part arrives at an angle nobody planned for. The claim is temporal. The medium is not. NVIDIA's glossary entry on physical AI defines the category as what "enables autonomous machines, such as robots in factories and robotaxis, to perceive, understand, and perform complex actions in the physical world", and every verb in that definition happens over time.

This is different from the problem a semiconductor company has. A chip is also invisible, but its argument compresses into numbers: latency, throughput, power per bit. A specialist reads the table and knows. Physical AI resists that compression, because the interesting figure is almost always conditional. Ninety-nine percent pick success is a different claim in a bin of identical parts than in a bin of mixed ones, and the condition is the part a table tends to lose. The Dex-Net 2.0 paper by Mahler et al., presented at Robotics: Science and Systems 2017, shows how much the condition carries: the same grasp planner reports a 93% success rate on eight known objects with adversarial geometry and 99% precision on 40 novel household objects, two different measures on two different object sets, all tested on singulated objects rather than a mixed bin.

So the spec sheet understates and the photograph misses entirely. What is left is motion.

What is actually invisible in an autonomous system?

The autonomy. Which is inconveniently the entire differentiator.

Put two machines side by side doing the same pick. One is running a taught path and will fail the moment the world moves. The other is perceiving, planning and deciding continuously. On camera they are indistinguishable. A buyer watching raw footage of your system has no way to tell your work from a teleoperated demo, and the more competent your engineering is, the more effortless and therefore the more ordinary it looks.

This is the specific job that motion graphics and 3D do in this category, and it is not decoration. Showing the depth map, the segmentation, the candidate grasps that were considered and rejected, the confidence falling as the light changes: that is the product. It is also the only version of the product a viewer can actually assess.

Most physical AI companies film the arm. The ones whose films work film the decision.

Why is the film usually the only demo?

Because you cannot ship one.

A software company hands over a sandbox and the buyer convinces themselves in an afternoon. A physical AI company cannot. The equivalent step is a paid pilot on the customer's floor, which costs both sides real money, takes months to schedule, and does not get approved by someone who is merely curious.

Everything before that pilot is media. The film is not marketing support for a sales conversation, it is the qualifying instrument that decides whether the sales conversation happens. B2B buyers now work through a long stretch of their decision before they contact anyone, and in this category the material they work through is the only demo available.

That reframes the budget question. A film that shortens the distance to a pilot, or that disqualifies a bad fit before either side spends a quarter on it, is not a communications cost.

When does 3D beat a camera?

3D beats a camera whenever the thing you need to show does not exist yet, cannot be opened or cannot be staged safely, which is more often than people expect and has nothing to do with budget.

A camera can only film what exists, where it exists, in the condition it is in. That rules out a great deal of what a physical AI company most needs to show:

• The fleet that is not deployed yet. You are selling two hundred units into a facility that currently holds four. The camera can film four.
• The mechanism inside the sealed housing. Cutaways, actuator paths, sensor placement and field of view. A camera would need the product destroyed, and often the interesting part is under an NDA anyway.
• The failure mode. Buyers ask what happens when it goes wrong, and the honest answer is the most persuasive thing you own. Staging a genuine failure repeatedly, safely, on camera is impractical. Simulating it is not.
• The site that has not been built. Greenfield deployments, new line layouts, throughput at a scale the customer has not reached.
• Scale of time. A cycle that takes nine hours, or a wear pattern that emerges over eighteen months.

None of those are a camera compromising. They are a camera being the wrong instrument. The category's own engineers already work this way. Waymo's February 2026 announcement of its World Model says it can simulate "exceedingly rare events" that are "almost impossible to capture at scale in reality", and NVIDIA's glossary describes digital twins used to test "fleet movements, layouts, and workflows" across a facility. The reverse holds too: when the machine exists and works, live action carries a credibility that a render cannot buy, because a render proves only that somebody can model. The choice is not aesthetic. It is about what each medium is able to prove.

What about hardware you are not allowed to show yet?

Hardware you cannot show yet is shown in 3D, and this is the case that catches companies out. It is common in physical AI because so much of the category is pre-launch, defence-adjacent or under customer confidentiality.

You cannot film a product that does not exist in final form, and frequently you cannot film the customer site at all. 3D becomes the only publishable representation, built from CAD and the spec rather than from a photograph. That has a second benefit worth naming: a model built from the engineering is correct in ways a shot of an early prototype is not, and it does not go out of date the week the enclosure changes.

We wrote separately about branding under NDA, and the pattern there holds here. The constraint is real and it is not a reason to publish nothing.

Does this replace the website and the deck?

No, and treating film as a substitute is the more expensive mistake.

The film earns attention and carries the sequence. The site has to survive the specialist who arrives afterwards wanting conditions, units, integration requirements and what happens at the edges. The deck has to survive a diligence call. They are different instruments answering different readers, and a company that makes a beautiful film sitting on a page that cannot answer a question has moved the failure rather than fixed it.

The order that works is the ordinary one: settle what you are claiming, then decide which medium can carry which part of it. Film is unusually load-bearing in this category. It is still not the whole argument.

What should a physical AI company make first?

One sequence, shown properly, of the thing you do that the alternatives cannot.

Not a company film. Not a montage of your hardware looking capable in flattering light, which is the default output and is interchangeable with every competitor's. One real task, in conditions you are willing to state, with the perception and decision layer made visible, and with the moment it nearly fails left in.

That last part is counterintuitive and it is the part that works. An engineer watching a system handle a case it clearly found difficult believes the system. An engineer watching a flawless loop assumes editing.

If you want the mechanics of how these get made, how a deep tech explainer video actually gets made covers the process, and 3D industrial animation covers when 3D is the wrong answer as well as the right one. Film production for deep tech companies covers which format fits which mechanism.

Book a 30-minute audit →