Robots are moving toward control systems powered by generative AI models. That shift creates a central safety question: if the software is probabilistic rather than fully predictable, how can companies prove the machine will behave safely around people?
Safeworld, a new company founded by Dr. Ding Zhao, Kyle Wong and Simo Rachidi, is built around that problem. The company is emerging from stealth with a seed round of more than $12 million and a plan to evaluate robotic control systems in simulation before real robots meet real workers.
Why generative AI changes the robot safety problem
Traditional software can often be checked against clearly defined rules. Generative AI systems are different. They can be powerful, but their behavior is harder to predict in every edge case, especially when they are controlling a physical robot.
Zhao, who directs the Safe AI lab at Carnegie Melon University, frames the issue as both technical and social. Robot builders must understand how to measure risk in probabilistic systems, but they also have to earn trust from the people and companies expected to use those systems.
That combination matters because robots do not operate only in clean demonstrations. They may be deployed in factories, construction sites and other places where people move unpredictably, visibility changes and each location has its own safety expectations.
Safeworld’s investors include Shine Capital and a16z Speedrun, which led the round, with additional investment from Box Group, Carnegie Mellon University Endowment, Innovation Endeavors and SV Angel. a16z Speedrun partner Jonathan Lai told TechCurnch that the safety standard should be built while robots are still being designed and deployed, not after household incidents involving kids have already happened.
What Safeworld tests in simulation
Safeworld’s core work is evaluating a robotic control system inside simulated environments populated with realistic human models. The company’s approach is similar in spirit to the testing challenge faced by companies like Tesla or Wayve, which must check how vehicles respond to unusual road situations.
Robots, Zhao argues, face an even more complicated problem. They operate in less structured environments than roads, and the standards for safe behavior may vary from one facility to another.
One example is a blind corner in a factory. Safeworld can build a digital version of that corner in a model like Genesis or MuJoCo, add a simulation of the robot being tested, run it with the robot’s real software and then try thousands of scenarios in which human models encounter the machine.
Those scenarios are designed to answer practical questions. How fast should the robot move near a blind corner? What stopping distance is needed? If a person is carrying boxes, will the robot detect that person in time?
Wong also pointed to tripping and falling as a scenario Safeworld tests in simulation. The point is straightforward: some safety cases are difficult, inefficient or unreasonable to recreate repeatedly with real people.
Why third-party validation may matter
Robot builders already use internal tools to test their systems. Safeworld’s founders believe an outside evaluator can still become important, both because of the company’s specialized focus and because third-party validation can help share safety knowledge across competitors.
That role becomes more relevant as robots move from controlled demos to broader deployment. Zhao said the concern is not the robot shown in a vacuum-like demo setting, but the robot used at scale by people who may never have operated a robot before.
For companies developing AI-driven machines, the safety bar is not only whether the system works most of the time. It is whether it can handle the messy cases: people in unexpected positions, people entering from awkward angles, people carrying objects, people falling and people behaving in ways the robot’s software did not see in a polished demonstration.
This is why simulation has appeal. It lets developers push a control system through many variations without waiting for each situation to occur in the real world. It also creates a way to compare how the robot behaves across repeated tests.
Gritt Robotics shows the real-world stakes
One early partner is Gritt Robotics. Vishal Dugar, the company’s CTO, is developing the AI brain for robots that currently help workers install photovoltaic panels at industrial-scale solar farms. The company also wants those robots to take on more complex construction tasks.
Gritt Robotics is working with Safeworld as it develops safety simulations. Dugar said systems like these are difficult to prove safe through formal math alone. In his view, the work has to be done empirically.
The reason is visible in the operating environment. Gritt Robotics’ machines work alongside human workers, so preventing a robotic arm from hitting a person is a central concern. To test that properly, the company has to consider many possible human behaviors and appearances.
Dugar described people kneeling, standing, tripping, falling, crouching and running. He also noted variation in clothes, size, shape, height, skin color and other appearance differences. For robot safety, those are not abstract concerns; they are the kinds of inputs a machine may have to interpret correctly on a work site.
An early company in an early market
Safeworld is still deciding what its product model should look like. The company is considering a platform for external users, a services based approach, or some combination that fits the market.
What is clearer is the problem the founders want to own. As generative AI becomes part of robotic control, robot makers need ways to evaluate unpredictable edge cases before deployment. They also need a credible way to show that work to customers, partners and the people who will stand near the machines.
Zaho expressed confidence that Safeworld can turn that need into a business, saying the company will probably be the first profitable company in the field because deployment requires someone to handle these safety situations.
The broader implication is simple: the robot industry’s next phase will not be judged only by what machines can do. It will also be judged by whether companies can show, in specific and repeatable ways, that those machines can work around people without creating unacceptable risk.