By Casey Forge
Put a generative AI model in charge of a robot, and you get something older factory code never was: a machine that can surprise you. So how do you prove it will not hit the person around the next blind corner?
The bet. Safeworld, coming out of stealth with a seed round of more than $12 million, says the answer is thousands of digital rehearsals — run before the robot ever shares a real floor with people.
Who is behind it. Ding Zhao, who directs the Safe AI Lab at Carnegie Mellon University (CMU), founded the company with veteran startup executive Kyle Wong and machine learning engineer Simo Rachidi. Shine Capital and a16z Speedrun led the round, with Box Group, the Carnegie Mellon University Endowment, Innovation Endeavors, and SV Angel also investing. The reporting so far comes from TechCrunch; we found no separate company announcement at filing.
A digital copy, the real software. Safeworld builds a digital version of a real spot — say, a blind corner in one particular factory. It drops in a simulated robot run by that robot’s actual control software, then plays out thousands of encounters with realistic human models: people carrying boxes, crouching, kneeling, running, even tripping and falling. The questions are simple: will the robot see the person, and can it stop in time?
Why generative AI changes the problem. Older robots follow predictable rules. Robots run by generative AI work on probabilities, so they do not always do the same thing twice. Zhao says the hard part has two halves: measuring the risk of a system like that, and earning the trust needed to use it. Robots also work in messy spaces, and each site has its own safety rules. a16z Speedrun partner Jonathan Lai told TechCrunch the time to build an industry safety standard is now, while robots are still being designed and rolled out — not after robots in homes start colliding with kids.
An outside referee, not just in-house tools. Robot makers already run their own simulations. Safeworld’s founders believe builders will also want a third party to check their work — if only to share safety lessons between competitors. Gritt Robotics, whose robots help workers install solar panels at large solar farms, is partnering with Safeworld on safety simulations. Its chief technology officer, Vishal Dugar, put it bluntly: you usually cannot prove these systems are safe with equations; it “necessarily has to be done empirically” — by testing against how people actually look and move.
The safety thread continues. The Agility Robotics and FORT Robotics agreement on a safety kit for the Digit 5 humanoid, signed October 1, was a similar bet: robots that work without a fence need a trust layer, not just a clever demo.
Ink, not pencil: this is a seed-stage launch reported by TechCrunch — not a certified industry standard or a finished product everyone must buy. The company is still deciding whether to sell a platform or a service. “Industry referee” is the founders’ ambition, not an established role yet.
Why regular people should care
Robots are leaving the demo cage. If walking, grabbing machines run by generative AI start sharing aisles, job sites, and one day homes, someone has to test the ugly edge cases first — blind corners, a worker who trips, different clothes, sizes, and postures. Outside testing is part of how cars and planes earned public trust. Robots may need the same, ideally before the first serious accident sets the rules.
What's next
What to watch. Whether Safeworld publishes its own announcement and early customer results, and whether other robot makers start treating this kind of testing as a must before deployment. For now, the bet is clear: more than $12 million to crash-test AI robots against simulated people in copies of real workplaces — while the industry is still young enough to set its own standard.