ROBBRIEF
Safeworld runs thousands of simulated humans at a robot's control software before it meets a real one, a verification layer the AI driven robot era will need built.
A Carnegie Mellon Safe AI Lab spinoff emerged from stealth on Monday with more than $12 million to build a verification layer for AI-driven robots, a step founders argue is overdue before such machines reach homes, hospitals, and factory floors.
Safeworld, founded by Dr. Ding Zhao, Kyle Wong, and Simo Rachidi, builds digital twins of operating sites in simulators like Genesis or MuJoCo, then runs the robot's actual control software against thousands of human-model scenarios. The goal is an empirical safety record for control systems that are probabilistic rather than rule-based, and whose behavior shifts with each new model update.
The company is betting this record, not a static certification, is what insurers, regulators, and factory safety officers will underwrite. Wong pointed to a near-term factory case: a robot must reliably stop at a blind corner when a person carrying boxes steps into its path, a test with no shared benchmark today.
Lead investors Shine Capital and a16z Speedrun, with Box Group, the CMU Endowment, Innovation Endeavors, and SV Angel, are funding the build. A16z partner Jonathan Lai said the work has to land "now, while robots are being designed and deployed," because household collisions would be "way too late."
The honest limits remain: probabilistic evals may not transfer cleanly to every real site, where safety expectations differ. Watch item: whether the bet becomes the underwriting standard before the first household incident forces one.