More than $12 million in seed funding has gone to a startup whose product is, in effect, a crash test for robots that decide for themselves. Safeworld came out of stealth on Monday with a round led by Shine Capital and a16z Speedrun, TechCrunch reported.
Other backers include Innovation Endeavors, SV Angel and the Carnegie Mellon University endowment. The link there is direct. Co-founder Ding Zhao directs the university’s Safe AI lab.
The problem is new. A conventional industrial robot follows fixed instructions and can be fenced off. One driven by a generative AI model behaves probabilistically, so the same situation can produce different movements.
Safeworld builds a digital copy of the place a robot will work, inserts realistic human models and runs thousands of scenarios. Co-founder Kyle Wong said the questions are concrete, such as what speed or stopping distance ensures the machine will not collide with a particular person.
People tripping and falling near a robot is a common test, and one nobody wants to stage for real. The simulations run on existing physics engines, including Genesis and MuJoCo.
The named partner is Gritt Robotics, which builds AI-driven machines for solar farm installation. Vishal Dugar, Gritt’s chief technology officer, said safety of this kind is very hard to prove formally with mathematics.
Zhao predicted Safeworld would probably be the first profitable company in its field. The founders have not yet decided whether to sell a platform or a service.
The asterisk is authority. A simulated pass means only what a regulator, insurer or customer decides it means, and the report names none of the first two. Whether simulated humans fall the way real ones do is the assumption the whole product rests on.
