Safeworld, a startup that tests robot software against computer-generated people, has come out of stealth with a seed round of more than $12 million. TechCrunch’s Tim Fernholz reported the launch on 5 October 2026. The pitch rests on a real tension in robotics: the newest control software is more flexible than the old kind, and also harder to predict.
Shine Capital and a16z Speedrun led the round. Four other backers joined them: SV Angel, Innovation Endeavors, Box Group, and the Carnegie Mellon University Endowment. The university link runs deeper than money. Dr. Ding Zhao, who directs the Safe AI lab at Carnegie Mellon, co-founded the company with Kyle Wong, described as a veteran startup executive, and Simo Rachidi, a machine learning engineer.
The problem is easy to state. Robot makers are handing more of the controls to generative models, the same family of technology behind chatbots. Older robot software followed rules an engineer could read and check line by line. A generative model learns its behaviour, so two similar situations can produce different reactions, and nobody can fully list them in advance.
Zhao frames the job as two parts. One is “probabilistic evals,” which in plain words means running enough tests to put a number on how often a system that behaves somewhat randomly will go wrong, so a risk can be priced and insured. The other, he says, is “the trust part.” In his telling, a company needs both before it can deploy a robot.
The method is a simulator. Safeworld builds a digital copy of a work site in a physics engine such as Genesis or MuJoCo, loads the customer’s real robot software into it, and then sends thousands of simulated people past the machine. Wong’s examples are practical: a worker carrying boxes who may or may not be detected, a blind corner where the robot’s stopping distance decides whether anyone gets hurt, and a person who trips and falls.
That last case explains the economics. Nobody wants to stage a hundred real falls in front of a moving robot arm. A simulation can stage a hundred thousand, with different body shapes, clothing, and postures. Vishal Dugar, the chief technology officer of Gritt Robotics, makes the same point from the customer’s side. His company builds AI for robots that help install solar panels, and he says systems like his cannot be proven safe with equations, so the proof has to come from testing. Gritt is partnering with Safeworld on its safety simulations, and the source does not describe it as a paying customer.
The comparison to self-driving cars is the founders’ own, and it cuts both ways. Car companies simulate crashes at enormous scale, yet Zhao argues robots are harder because they work in messy indoor spaces, and because every facility sets its own safety rules.
Here the evidence ends. TechCrunch reports no customer results, no data showing that a robot which passes Safeworld’s simulations behaves safely around real people, and no decision yet on whether the product will be a platform or a consulting service. A seed round measures investor confidence, not demonstrated outcomes. Robot makers also already run internal simulators, so Safeworld must convince them that an outside tester adds something their own tools do not. The founders’ answer is that competitors may want a neutral party to vouch for safety cases they can all share.
If that neutral referee idea holds, safety testing could become a purchase line on a robot maker’s budget, much as security audits did for software. Anyone buying or deploying humanoid or warehouse robots should ask vendors now whether their safety claims come from tests they ran on themselves or tests an outsider ran for them.
Reported by Tim Fernholz for TechCrunch, published 5 October 2026.
