Safeworld robot safety testing is the pitch behind a start-up that came out of stealth on 5 October 2026 with a seed round of more than $12 million. The company builds simulations full of realistic digital humans, then runs a robot’s real control software through thousands of scenarios to see whether it would hurt anyone. Its founders include Ding Zhao, director of the Safe AI Lab at Carnegie Mellon University.
The timing is not an accident. Robot makers are handing control to generative AI models, which learn behaviour from data instead of following hand-written rules. As TechCrunch put it, that architecture “isn’t predictable the way traditional algorithms are”. A robot that cannot be proven safe on paper has to be shown safe some other way.
Below we explain what Safeworld announced, why GenAI robots need a new kind of safety test, how Safeworld robot safety simulations work, what a solar-farm robot maker hopes to get from them, the case for and against a third-party tester, the standards and laws now catching up, and what businesses deploying robots should ask for.
Table of contents
- What Safeworld Announced
- Why GenAI Robots Need a Different Kind of Safety Test
- How Safeworld Robot Safety Testing Works
- The Blind Corner Problem: Why Stopping Distance Matters
- Gritt Robotics and the Solar Farm Test Case
- Can Safeworld Convince People? The Case For and Against
- Safeworld Robot Safety vs Field Trials and In-House Simulation
- Robot Safety Rules Are Catching Up
- What Safeworld Robot Safety Means for Businesses Deploying Robots
- What to Watch Next for Safeworld Robot Safety
- Safeworld Robot Safety: Frequently Asked Questions
- References
What Safeworld Announced
Safeworld, which styles itself SafeWorld in its own announcement, is based in Palo Alto and describes itself as “an AI lab building robot safety simulation technologies”. The Safeworld robot safety business sells to robot makers and to the companies that deploy their machines.
The round and the investors
| Item | Detail |
|---|---|
| Round | Seed, $12.2 million (TechCrunch: “more than $12 million”) |
| Co-leads | Shine Capital and a16z Speedrun |
| Other investors | Box Group, Carnegie Mellon University Endowment, Innovation Endeavors, SV Angel, plus angels |
| Angels from | NVIDIA, Google DeepMind, Waymo, Meta and DoorDash, according to citybiz |
| Early customers | Pilots with Fortune 50 firms, including automotive manufacturers, medical device makers and warehouse automation companies |
| Named partners | Gritt Robotics and Anyware Robotics |
The founders
Ding Zhao leads the Safe AI Lab at Carnegie Mellon, is a former Google DeepMind researcher, holds a National Science Foundation CAREER Award and, according to the company, has spent more than 17 years on the safety of autonomous systems. Chief executive Kyle Wong founded Pixlee, a content platform used by more than 1,000 brands that was later acquired, and most recently ran StartX, Stanford’s start-up accelerator. Simo Rachidi was a principal security and machine learning engineer at Salesforce Einstein.
The promise
“As robotics moves from impressive demos to everyday deployment, safety becomes a prerequisite for adoption,” Wong said in the launch announcement. The Safeworld robot safety pitch is that if autonomy is going to scale to billions of machines, safety validation must scale with it.
Why GenAI Robots Need a Different Kind of Safety Test
Factory robots of the past were deterministic. They repeated programmed movements inside fenced cells, so engineers could reason about every motion. The robots now arriving in warehouses, hospitals and construction sites are steered by AI models that respond to what they see.
Probabilistic, not programmed
“The safety challenge that we’re talking about is a combination of, one, really advanced generative AI probabilistic evals — how do you underwrite the risk of a probabilistic system?” Zhao told TechCrunch. A model that usually behaves well can still do something unexpected in a situation it never saw in training. Safeworld robot safety testing exists to find those situations before a person does.
The trust problem
Zhao named a second difficulty: “The second part that’s really hard is the trust part, and you need both to deploy a robot.” Passing internal tests is not the same as persuading a factory safety officer, an insurer, a regulator or a worker standing next to the machine. That is the question in TechCrunch’s headline, and it is a commercial problem as much as a technical one.
Cages and speed limits
Safeworld’s announcement says that, faced with this uncertainty, deployers “resort to physical cages and speed limits, throttling robot productivity and human collaboration”. A robot that has to slow to a crawl whenever a person might be near loses much of its value. Better evidence of safe behaviour, which is what Safeworld robot safety testing sells, is what would let operators loosen those limits.
How Safeworld Robot Safety Testing Works
The Safeworld robot safety method borrows from the self-driving car industry, where companies such as Tesla and Wayve test how vehicles react to surprising events. Zhao argues robots are harder: they work in unstructured spaces, and each site has its own safety rules.
Step 1: build the scene
A Safeworld robot safety test starts with a digital version of the place the robot will work, such as a particular blind corner in a particular factory, in a physics simulator such as Genesis or MuJoCo. The announcement says teams can build scenarios “right in the browser from past incidents, safety standards, and robot logs, with no simulation expertise required”, using natural-language descriptions.
Step 2: drive the robot with its real software
A simulated version of the robot is dropped into the scene, controlled by its actual software rather than a simplified stand-in. That matters for Safeworld robot safety results, because the risk lies in how the real AI model behaves, not in an idealised version of it.
Step 3: run thousands of human encounters
The simulation then runs thousands of variations in which digital humans meet the robot. “Tripping and falling is also a good example of something that we do a lot of testing with the simulation,” Wong said. “Otherwise, you would have to go and trip and fall for the robot, which is like a hard thing to be doing all the time.” The company says its human models move realistically and react to the robot.
Step 4: rerun after every change
Because every software update or new site can introduce new risks, Safeworld robot safety tests are meant to be rerun continuously, not once before launch. The result, the company says, is “a safety record that engineering, safety, and operations leaders can stand behind”.
Scenarios named so far
| Scenario | The question it answers | Who raised it |
|---|---|---|
| Blind corner in a factory | What speed or stopping distance prevents a collision? | Kyle Wong |
| Worker carrying boxes | Will the robot still detect a partly hidden person? | Kyle Wong |
| Person trips and falls | Does the robot react safely to a sudden fall? | Kyle Wong |
| Kneeling, crouching or running workers | Does detection hold across body positions? | Vishal Dugar, Gritt |
| Varied clothes, size, height and skin colour | Does detection hold across every kind of person? | Vishal Dugar, Gritt |
| Robots near untrained people | Is it safe at scale, not just in a demo? | Ding Zhao |
The Blind Corner Problem: Why Stopping Distance Matters
Wong’s blind-corner example is worth working through, because it shows why small changes in a robot’s behaviour matter.
The arithmetic of a delay
“What is the speed or what is the stopping distance that you need to make sure that this robot will not collide with a particular human?” Wong asked. Before a robot can brake, it has to detect the person and decide to stop. During that delay it keeps moving at full speed. Distance covered equals speed multiplied by time, so a 0.5-second delay at 1.5 metres per second means 0.75 metres travelled before braking even begins. The chart applies that sum at four speeds; the figures are an illustration, not Safeworld data.
Why simulation helps here
Doubling the speed doubles the distance covered before braking, and a model that takes longer to recognise a person carrying boxes adds more. Measuring those delays across thousands of simulated encounters, at every corner of a real site, is exactly what physical trials cannot do cheaply. It is the core of the Safeworld robot safety offer.
Gritt Robotics and the Solar Farm Test Case
The first named Safeworld robot safety partner shows where the problem bites. Gritt Robotics, which TechCrunch reported in July raised $34 million, builds robots that help workers install photovoltaic panels on industrial-scale solar farms and aims to take on harder construction tasks.
Why maths alone cannot prove safety
Vishal Dugar, Gritt’s chief technology officer, is developing the AI that controls those robots. “The difficulty with most of our systems is it’s very hard to formally prove it by doing some math, writing some equations, and saying yeah, the system is verified to be safe,” Dugar said. “It necessarily has to be done empirically.” Gritt’s robotic arms work alongside people, so not hitting them is the first requirement.
Humans come in every shape
“Humans have many kinds of appearances,” Dugar said. “Their bodies can be in different configurations. They could be kneeling, standing. They could be tripping and falling potentially. They could be crouching. They could be running.” Dugar added that robots must also cope with variation in “clothes, size, shape, height, skin color, everything else.” That is a fairness question as well as a safety one: a detector that works less well on some people puts them at greater risk, and Safeworld robot safety scenarios will need to cover that variation explicitly.
What a partner gets
For a company like Gritt, Safeworld robot safety simulations offer a way to collect evidence across far more situations than a field trial on a live solar site could safely stage. Anyware Robotics chief executive Thomas Tang said in the announcement that simulation tools give the company “a scalable way to test challenging scenarios” and better prepare its robots for real-world deployment.
Can Safeworld Convince People? The Case For and Against
The headline question is whether Safeworld robot safety evidence can win trust, and the honest answer is that the company has not yet had to earn it at scale.
The case for an independent tester
Safeworld’s founders argue that robot makers will want a third party to validate their work, not least so safety cases can be shared between competitors. Cars offer a precedent: independent crash-test programmes changed how vehicles were designed because buyers could compare published results. Zhao is bullish about the business: “We’ll probably be the first profitable company in this field. Because if anyone wants to deploy, they need to pay us to handle the situation.”
The case against
TechCrunch notes “definite similarities” between Safeworld’s platform and tools robot builders already use internally. Simulations are only as good as their models of people and physics, and a gap between simulated and real behaviour is a known problem in robotics. Safeworld robot safety results will also need someone to check the checker. The company is still deciding whether to sell a platform for customers to run themselves or a services-based offering.
Why trust is fragile
Concern about humanoid strength is no longer hypothetical. In November 2025 Robert Gruendel, Figure AI’s former principal robotic safety engineer, sued the company in federal court in California, alleging Gruendel was fired after warning that its robots were powerful enough to fracture a human skull and that one had cut a quarter-inch gash in a steel fridge door. Figure said the dismissal was for poor performance and that it would discredit the claims.
a16z Speedrun partner Jonathan Lai put the investor’s view bluntly: “By the time you have robots in households—colliding with kids—and causing safety incidents, that’s way too late.” Cases like Figure’s are why Safeworld robot safety evidence would need to be independent to persuade anyone.
Safeworld Robot Safety vs Field Trials and In-House Simulation
Robot makers already have ways to test their machines. The table sets the Safeworld robot safety approach against the alternatives, based on what the company and its partners have said.
| Approach | Strength | Weakness |
|---|---|---|
| Physical field trials | Real robots, real physics, real sites | Slow and costly, and falls or near-misses cannot be staged safely at scale |
| Cages and speed limits | Simple and well understood | Throttle productivity and keep people and robots apart |
| Formal verification | Mathematical guarantees | Hard to apply to learned controllers, as Gritt’s CTO says |
| In-house simulation | Engineers know the robot best | Makers mark their own homework, and results are rarely shared |
| Safeworld robot safety simulation | Thousands of scenarios, run by a third party, rerun after every update | Only as good as its human and physics models; not yet proven at scale |
Where Safeworld robot safety fits
The likely answer is a mix. Field trials will still be needed to confirm that simulated results match reality, and cages will stay for the most dangerous tasks. What Safeworld robot safety testing adds is breadth: rare situations, such as a worker tripping into a robot’s path at a blind corner, can be run thousands of times without anyone getting hurt. It also adds an outside view, which in-house simulation cannot offer however good it is.
The sim-to-real question
Robotics researchers have long struggled with the gap between how a machine behaves in simulation and how it behaves on a real floor. Safeworld says its digital humans move realistically and react to the robot, but it has not published how those human models were checked against real human motion. For Safeworld robot safety results to carry weight with insurers, notified bodies and site safety teams, that validation will need to be documented in detail.
Robot Safety Rules Are Catching Up
Safeworld robot safety tooling lands as standards bodies and regulators start writing rules for AI-driven machines.
The standards and laws that matter
| Rule | Status | Why it matters for GenAI robots |
|---|---|---|
| ISO 10218-1 and 10218-2:2025 | Published February 2025 | First major revision of industrial robot safety since 2011; collaborative robot rules from ISO/TS 15066 folded into Part 2 |
| ISO 25785-1 | Committee draft | First safety standard for mobile robots that must actively balance, such as humanoids and quadrupeds |
| EU Machinery Regulation 2023/1230 | Applies from 20 January 2027 | Self-learning software that performs safety functions needs a third-party check by a notified body |
| EU AI Act, Annex I products | 2 August 2028, after the AI Omnibus | AI safety components of machinery become high-risk AI systems |
| US workplace safety law | No robot-specific OSHA standard | Employers rely on general duties and industry standards |
Europe’s third-party requirement
The EU rule is the most important for the Safeworld robot safety business model. From 20 January 2027, machinery whose safety functions depend on “self-evolving behaviour using machine learning” must go through a notified body before it can carry a CE mark in the EU. That is a legal demand for exactly the kind of outside evidence Safeworld wants to supply, although the notified bodies themselves, not a start-up, will sign off.
Humanoids are still ahead of the standards
ISO 25785-1 is still at committee draft stage, so humanoid robots entering warehouses today have no finished standard written for them. Until it is published, buyers will lean on general risk assessments and on whatever evidence vendors can produce, which is the gap Safeworld robot safety testing is aimed at.
What the record says
Industrial robots have been dangerous before. A NIOSH study of US fatality data found 41 robot-related worker deaths between 1992 and 2017. Most involved fixed robots, and most were a robot striking a worker, often during maintenance. The chart shows the shares the study reported.
Those deaths came from caged, programmed machines. Mobile robots driven by AI models will be closer to people more of the time, which is why the industry is looking for better ways to test before deployment.
What Safeworld Robot Safety Means for Businesses Deploying Robots
Most businesses will meet this issue as buyers of robots, not builders. Safeworld robot safety testing is aimed at both groups, but the questions for buyers are the same whoever does the testing.
Questions to ask a robot vendor
- Which scenarios has the robot been tested against, and were they specific to our site?
- Were tests run with the robot’s real control software, and are they rerun after each update?
- How does detection perform across body positions, sizes, clothing and skin tones?
- Which standards does the robot meet today, and what is the plan for the EU Machinery Regulation?
- Who outside the company has reviewed the safety case?
Treat updates as a safety event
A robot whose behaviour comes from an AI model can change with a software update. Make updates part of your change control, with retesting before rollout, and treat the update channel as a cybersecurity risk, because a tampered update is a safety hazard as well as a data one. Our guide to robot fleet management software covers how to control versions across a fleet.
Keep people in the design
Safety in shared spaces depends on how people and robots coordinate, not only on the robot. Our report on Destro AI’s work getting robots and humans on the same page looks at that side of the problem, and our piece on how virtual playgrounds help robots get training data explains why simulation has become central to robotics.
What to Watch Next for Safeworld Robot Safety
Three signals will show whether Safeworld can turn a strong research pedigree into the trust it needs.
A published safety case
The most persuasive step would be a customer agreeing to publish what a Safeworld robot safety campaign found, including failures. Evidence that people can read will do more than funding announcements.
The product decision
Safeworld says it is still choosing between a self-serve platform and services. A platform scales faster; services may suit Fortune 50 buyers who want an independent report with a name attached.
Standards and notified bodies
Watch whether simulation evidence is accepted in ISO 25785-1 and by EU notified bodies under the Machinery Regulation. If regulators accept simulated testing as part of a safety case, Safeworld robot safety tools could become part of the approval process rather than an optional extra. We looked at the same tension in safety-critical software in our coverage of building AI when failure is not an option.
Safeworld Robot Safety: Frequently Asked Questions
What is Safeworld?
Safeworld is a Palo Alto start-up, co-founded by Carnegie Mellon’s Ding Zhao, Kyle Wong and Simo Rachidi. Its Safeworld robot safety platform builds simulations to test whether AI-controlled robots will behave safely around people.
How much has Safeworld raised?
It raised a $12.2 million seed round, co-led by Shine Capital and a16z Speedrun, announced on 5 October 2026.
How does Safeworld robot safety testing work?
It rebuilds a real workplace in a physics simulator such as Genesis or MuJoCo, runs the robot’s real software inside it, and tests thousands of encounters with realistic digital humans, including people tripping, falling or carrying boxes.
Why can’t GenAI robots be tested the old way?
Robots steered by generative AI models behave probabilistically, so their actions cannot be fully proven with equations. Their safety has to be shown empirically, across many scenarios.
Is simulated testing required by law?
Not as such. From 20 January 2027, the EU Machinery Regulation requires third-party conformity assessment for machinery whose safety functions rely on machine learning, and simulation is one way to produce the evidence.
References
Can Safeworld convince people that gen AI robots won’t hurt them? (TechCrunch)
Introducing SafeWorld: an AI lab building systems to improve robot safety (Robotics Tomorrow)
SafeWorld emerges from stealth with $12.2M (citybiz)
Gritt exits stealth with $34 million for robots to build solar plants (TechCrunch)
Updated ISO 10218: frequently asked questions (A3)
ISO/CD 25785-1 (International Organization for Standardization)
Type examination of machines that use AI for safety functions (Danish Technological Institute)
Study of robot-related worker deaths highlights safety challenges (Safety+Health)
