Podcast

A Warehouse Is Not the World

AI
Automation
Logistics
Deployment

Why physical AI is safe in structured environments and still unproven in the unstructured world we actually live in.

1:35

Self-Driving Cars Are Pushovers “Oh, they just get out of your way."

Why do humans instinctively bully autonomous cars? Because the car flinches first. A look at how self-driving cars behave in unstructured traffic, and what happens when your car can circle the block instead of parking.
11:18

Fitbits for Cows “Maybe the funniest, weirdest project I’ve ever worked on.”

What does the USDA want with edge machine learning and a herd of cattle? More than you’d guess. The strangest data problem the team ever took on, and why sparse, messy real-world inputs break traditional software.
17:42

Box Yes, Human No “A robot is allowed to touch a box, but not a human.”

When does a warehouse robot become a public one? The second it leaves the warehouse. The crux of the safety problem and why people are instinctively great at the nuance that machines still can’t grasp.
28:50

The Box That Wiggled “It’s okay for a robot to pick up a box. But what if my kid’s in it?”

A parent’s instinct catches what a robot can’t: the box is heavier than it should be, and it just wiggled. A grounded gut-check on how far home robotics really is, and why you should test the vision models yourself before trusting the hype.
41:11

When Machines Seem Alive “Truly alive or not, when does it become a weird social problem?”

We blew past the Turing test years ago. So what happens when a home robot looks human, acts human, and starts goal-seeking for your approval? A candid debate on machine consciousness, misalignment, and the strange social questions we may be creating before we are ready.

A robot is allowed to touch a box, but it’s not allowed to touch a human.

Jami Friedman, LeafLabs

Structured vs. Unstructured: The Real Safety Problem in Physical AI

Our conversation with LeafLabs kept coming back to one question: what does it actually take to make physical AI safe? The answer turns on a distinction most product roadmaps gloss over. There are structured environments, where the world is controlled and predictable, and there are unstructured ones, where it is not. A robot can look flawless in the first and be genuinely dangerous in the second, and the gap between the two is where almost all of the hard engineering lives.

In a warehouse or a lab, you can define the space. On a sidewalk, in a hospital corridor, or in a living room, you cannot. As robots move out of controlled settings and into the places people actually live, the safety problem does not grow gradually. It changes character entirely.

Read more Show less

Structured vs. Unstructured is Really a Spectrum of Predictability

A conveyor belt is structured because you already know roughly 95 percent of what is coming and where. The last 5 percent, like guessing the weight of a box before you lift it, is an inference problem humans solve from a hundred small cues without thinking. Robots do not. Watch a robot grab a cup it has never weighed, and you can see the guess happen. Move that same robot into an unstructured space, and the share you cannot predict climbs from 5 percent to most of the picture.

In Structured Environments, You Buy Safety with Control

Warehouses stay safe through fences, floor tape, beacon vests that slow nearby robots, training, and the blunt option of removing anyone who ignores the rules. Even there, the machine still has to operate with enough freedom to do its job without swinging an arm into a trailer. None of those controls transfer to the open, unstructured world. You cannot fence a sidewalk or train the public, so in unstructured settings, the safety parameters have to come from the system itself.

Safety is Not One Rule; It is Endless Nuance

A useful robot has to be allowed to pick things up. That is the job. But it also has to be able to distinguish between a box it can move and one it must not, the one that is heavier than expected because a child is hiding inside. A robot can touch a box; it cannot touch a human. Writing that distinction into policy broadly enough to cover every real case, while still letting the machine work, is one of the hardest open problems in robotics. People parse it effortlessly. Today’s models are leaps away.

The Technology is Not There Yet, and Pretending Otherwise is the Risk

We do not have true world models. Vision models can describe a scene impressively, and a language model will happily generate a 3D object whose code runs but makes no spatial sense. Closing the gap will take a different and arguably new class of model, plus two things we mostly lack: a way to prove that certain unsafe actions are structurally impossible, such as a hard swing while a human is in range, and a way to audit behavior in real time and stop fast when it deviates.

The Engineering Answer is the Same as it Always Was: Shrink the AI Surface

Bound the problem. Find the one piece that genuinely needs non-deterministic behavior, solve that with AI, and codify everything else. Flexible inputs, deterministic outputs. The home helper of the near future may not even be a humanoid. A ceiling-mounted arm, or a set of appliances that hand off to each other, can be safer and more capable for most tasks, even though the market keeps asking for something that looks like us.

Whatever you Build - Build it as a Good Steward of Humanity

Hardware decisions are not software decisions. The hardware you choose for a device is a choice you live with for five, ten, or, in agriculture, possibly up to thirty years, and nothing is more permanent than a temporary solution. Rightsize it, and keep the AI contained, wrap it in deterministic guardrails, and be honest about what the system can and cannot prove about its own safety.

If you are deciding whether AI belongs in a physical product, the real question is not whether you can add it. It is whether the environment is structured enough to make it safe, and whether you can live with the choice for a decade. That is the conversation worth having before the crate shows up. If you are working through that decision, we can help you test it before you commit.

Frequently Asked Questions

Frequently asked questions

What is the difference between a structured and an unstructured environment?

A structured environment is one you can mostly predict and control, like a conveyor belt, a fenced warehouse cell, or a lab, where you might know 95 percent of what will happen. An unstructured environment, like a home, a sidewalk, or a hospital hallway, is full of changes you cannot label in advance. The less predictable the space, the more inference the robot has to do, and inference is exactly where today’s machines are weakest.

Why does safety get so much harder in unstructured environments?

In structured settings, you manage risk with fences, floor tape, beacons, training, and the ability to remove people who break the rules. None of that works around the general public. Safety has to come from the system itself, which must distinguish, instantly, between what it may and may not act on, like a box it can lift versus a human it must never touch.

Is today’s AI good enough to make home and public robots safe?

Not yet. We lack true world models, the ability to prove that unsafe actions are impossible, and reliable real-time auditing that can stop a machine fast when it deviates. Vision-language-action models are impressive but spatially unreliable. The responsible approach is to keep the AI surface small and wrap it in deterministic guardrails rather than trusting a general-purpose model to handle everything.

If not a humanoid, what should a home robot look like?

Often something simpler and safer, like a ceiling-mounted arm or connected appliances that hand tasks to each other, because the human shape is not actually optimized for most household work. The counter-pressure is that people trust and demand machines that look like them, so the answer is genuinely contested.

About the Guest

Jeff Ciesielski

Director of Research

Ciesielski went to LeafLabs having spent the past decade or so of his career on a wide variety of software and hardware projects ranging from race car telemetry systems, to accounting software for commercial fisheries, to most recently bending a fleet of autonomous manufacturing robots to his will using Haskell. When he isn’t working, he enjoys spending time with his wife, tinkering on his never ending race car project, and tending a small pride of house cats.

Jami Friedman

Executive Vice President

Friedman leads operations, program management, and partnerships at LeafLabs, overseeing client engagements and the systems that keep complex projects on track. She brings over a decade of experience managing hardware development and machine learning programs. Outside of work, she’s either cooking for a crowd or hunting for mid-century furniture at estate sales. (University of Massachusetts Amherst, BBA Supply Chain & Operations Management ’13)

A Warehouse Is Not the World

Self-Driving Cars Are Pushovers “Oh, they just get out of your way."

Fitbits for Cows “Maybe the funniest, weirdest project I’ve ever worked on.”

Box Yes, Human No “A robot is allowed to touch a box, but not a human.”

The Box That Wiggled “It’s okay for a robot to pick up a box. But what if my kid’s in it?”

When Machines Seem Alive “Truly alive or not, when does it become a weird social problem?”