How you check that a robot is safe around people: the startup SafeWorld and $12.2 million for simulating rare situations

Short answer: on 8 October 2026 The Robot Report reported that, that week, the startup SafeWorld had emerged from stealth with $12.2 million in seed funding. The company builds a simulation in which a robot is tested not on whether it can do the job, but on how it behaves next to a person in a rare, dangerous or unexpected situation. Scenarios, by the company's description, are built in the browser from past incidents, safety standards and the robot's own logs, after which the robot is run through thousands of variations in which human motion in the model reacts to the robot. The reason the founders give: with the arrival of AI, robots were allowed out of the cage for the first time and now share space with people, while there is almost no data about such rare situations. Below: what exactly was announced, how this differs from ordinary simulation, and what to take away if you buy robots rather than build them.

What happened

SafeWorld emerged from stealth that week and announced $12.2 million in seed funding: that is how it is described in the publication of 8 October 2026. The round was co-led by Shine Capital and a16z Speedrun, with participation from Box Group, the Carnegie Mellon University Endowment, Innovation Endeavors, SV Angel and other investors.

The company is based in Oakland, California. According to its CEO Kyle Wong, it has customers in industry, manufacturing, logistics, construction and even some home use cases. The platform, by the company's statement, works with a range of robot embodiments: robot arms, humanoids, mobile robots and others.

Everything said below about how the platform is built and what it is good for comes from the company itself and its customer, in an interview with the outlet. I have not seen it in operation, and there are no independent checks in the source.

Why safety testing became a separate task

The most useful part of this material is the explanation of why a separate instrument was needed at all. SafeWorld co-founder Ding Zhao, director of the Safe AI Lab at Carnegie Mellon University, puts it this way: robots have been in use for maybe a hundred years, but with the arrival of AI robots were for the first time allowed out of the cage and to share the same space with humans.

The cage here is not a figure of speech. A classic industrial robot is fenced off, and the fence is one of the measures that keeps people away from its working area. When a robot works next to people without a fence, another task is added to the usual measures: you also have to answer for how the machine behaves while doing it.

Hence the difficulty Wong names: these are safety scenarios that have not existed yet in the world. Training for situations with no wide dataset behind them is what he calls a big challenge.

Wong's second thought is about the division of labour: robot manufacturers, of course, are responsible for safety themselves; the question is only whether they should build it from scratch every single time. In his view, that makes no sense.

How this differs from ordinary simulation

Simulation has long been used in robotics: a robot is run in a virtual environment so as not to break hardware. Wong says that with safety it breaks down in the detail: first you have to work out what even counts as the testing criteria. There are standards, he says, but they are not all complete, and on top of that come different kinds of risk assessment.

How the work is arranged, by the company's description:

• scenarios are built right in the browser from three sources: past incidents, safety standards and the robot's own logs, and, as the company states, no simulation expertise is required for this;
• then the robot is run through thousands of variations of that scenario, with human motion in the model that is realistic and reacts to the robot;
• behaviour is measured in terms of safety.

And a separate point worth highlighting: since every software update and every new environment can introduce new risks, such tests, the company says, can be rerun continuously rather than once before launch. SafeWorld calls the result documented test results, which engineering, safety and operations leaders can stand behind. The outlet presents this specifically as a claim by the company. There is no talk of official certification in the source.

When it is time to think about this

Wong gives his own rule of thumb on timing, softening it with "probably" and "maybe": if your deployment is less than a year away, you probably need to start thinking about safety; if it is further off, it is maybe a little too early. The source does not explain why.

In the same place he divides companies into two groups. Some bring AI to the robots: like Caterpillar, they already have sophisticated hardware working in the real world, and for them safety matters from day one. The second group he calls "bring robots to AI", and how exactly it is arranged is not covered in the article.

He himself works, by his own account, with about half a dozen companies, two-thirds of which are publicly traded. That is a short list, and it describes the current state of a startup rather than how widespread the approach is.

Why a third party is involved here

The argument both founders make deserves its own paragraph, because it applies far beyond robots. Wong says that safety has an engineering part and a trust part, and that you need both to deploy. The trust part, in his words, is about working with the people who take the decision, and the question of who actually signs off.

Zhao adds: this is a deployment problem, so a separate company is needed to win trust not only for itself but for the industry as a whole. Both are interested parties, of course: what they sell is precisely such a third party.

Among customers, the outlet quotes Anyware Robotics CEO Thomas Tang: in his words, safety has been foundational to the company from day one, and tools like this give a scalable way to test challenging scenarios and better prepare robots for real-world deployment.

What to take away if you are buying a robot

The list of questions below can be used without buying the platform at all: it is drawn from how its creators approach testing. It is better to ask them in writing, before paying for a robot. If the supplier cannot confirm safety and there is no way to run a check in your own conditions, there is one conclusion: the robot is not admitted to work next to people until the risks have been assessed and confirmations obtained.

1. What confirms the safety of this particular model in your setting. A robot in a warehouse with forklifts and the same robot in a workshop have different neighbours.
   • Who does it: you or your technical specialist.
   • How to check: you have a document or a letter naming specific tested situations, rather than general words about compliance.
2. What happens after a software update. A robot's behaviour is set by software, and software changes.
   • Who does it: you, when signing the contract.
   • How to check: the contract or the supplier's letter states whether tests are rerun after updates and who is responsible for that.
3. Which rare situations were tested. A person crouched to tie a shoelace, dropped a box, came round a corner, two people walked towards each other at the same time.
   • Who does it: you together with whoever works next to the robot.
   • How to check: draw up your own list of five such cases from your site and ask about each one whether it was tested. Keep the answers.
4. Who signs off on the launch. The argument about the trust part works at your end too: a decision needs a person behind it, not general agreement.
   • Who does it: the manager.
   • How to check: an order or minutes name the person responsible for admitting the robot to work next to people, with a date.
5. How the robot behaves on failure. It is worth asking separately not about normal operation but about loss of connection, a flat battery and an emergency stop.
   • Who does it: a technical specialist on your side.
   • How to check: the supplier has described the behaviour in these three cases in writing. Testing this at acceptance is only possible under a safe procedure agreed with the supplier, carried out by a competent specialist, and only when there are no people in the danger zone.

Where the data used to train robots comes from in the first place, I went through in the article Why startups pay people to record everyday chores. How industry gradually gives automation more independence is here: AI in manufacturing.

What not to expect from this news

• That simulation replaces trials. The source says it gives a way to test how robots behave around people in rare, dangerous and unexpected situations without putting anyone at risk. It says nothing about replacing real trials.
• That testing makes a robot safe. Testing gives a description of behaviour in the situations that were tested. Untested situations remain untested.
• That this is an industry standard. The company has only just emerged from stealth and works, by its CEO's account, with about half a dozen companies.
• An independent assessment of the platform. There is none in the source: there are the founders' words, a customer's words and a report of a funding round.

In summary

In its publication of 8 October 2026, The Robot Report reported that the startup SafeWorld had announced $12.2 million in seed funding for a simulation in which the behaviour of robots around people in rare and dangerous situations is tested. The founders' argument: AI brought robots out of the cage into shared space with people, while there is almost no data about such situations, so it has to be created in a model. Two thoughts are worth keeping separately. First: a robot's behaviour is set by software, and software changes, so the question "are tests rerun after updates" is one to put to the supplier. Second: safety has an engineering part and a question of who signs off. If you are buying a robot, both thoughts turn into clauses in a contract.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources

• The Robot Report, 08.10.2026: SafeWorld emerges from stealth to build, deploy robot safety simulation technologies. https://www.therobotreport.com/safeworld-emerges-stealth-build-deploy-robot-safety-simulation-technologies/