Can Safeworld convince people that gen AI robots won’t hurt them?

The rush to put generative AI in robots has created a new problem: the software is less predictable than the old rule-based stuff. Safeworld thinks that’s exactly where a business can be built. The startup just came out of stealth with more than $12 million in seed funding, backed by Shine Capital and a16z Speedrun, plus Box Group, the Carnegie Mellon University Endowment, Innovation Endeavors and SV Angel.

The company was founded by Dr. Ding Zhao, who leads the Safe AI lab at Carnegie Mellon University, along with Kyle Wong and Simo Rachidi. Zhao has spent years on this safety problem, and he sees two separate hurdles. One is technical: how do you judge the risk of a probabilistic system? The other is social: how do you get people to trust it enough to deploy it?

Safeworld’s pitch is that robotics needs an outside referee. The company wants to test robot control systems in simulations filled with realistic human models, using environments built in tools like Genesis or MuJoCo. The idea is to recreate nasty little moments that matter in the real world: a blind corner in a factory, someone carrying boxes, a person tripping or falling. Then the robot, running its actual software, gets thrown into thousands of scenarios.

That matters because robot makers already do some of this internally, but Safeworld’s founders think a third party will be useful anyway, especially if safety cases need to be shared across competitors. Jonathan Lai of a16z Speedrun called the timing right, before robots start showing up in homes and causing incidents with kids. Zhao’s own warning is simpler: the hard part is not the demo robot. It’s the one deployed at scale, with people who have never operated one before.

The company is already working with Gritt Robotics, whose systems help workers install photovoltaic panels at industrial-scale solar farms and are meant to move into more complex construction jobs. Gritt’s CTO, Vishal Dugar, says the safety case can’t be proven neatly with math alone; it has to be tested in practice. That means accounting for crouching, kneeling, running, falling, different body types, different clothing and all the messy variation that makes humans such a pain for robots.

Safeworld hasn’t settled on its business model yet. It could become a platform for outside users or a services shop. But Zhao is already calling the shot: if companies want to deploy robots, they’ll need someone to handle the safety burden, and they’ll have to pay for it.