Safeworld Raises $12.2 Million to Find Out What Generative-AI Robots Do Before You Deploy Them

Safeworld, a Carnegie Mellon University spinout that safety-tests robots driven by generative AI, came out of stealth on October 5 with a $12.2 million seed round co-led by Shine Capital and Andreessen Horowitz's Speedrun program. Box Group, the Carnegie Mellon University Endowment, Innovation Endeavors and SV Angel also invested, alongside angels from Nvidia, Google DeepMind, Waymo, Meta and DoorDash.

The problem Safeworld targets is structural. Robots controlled by generative AI models behave probabilistically — the same policy can produce different actions on different runs — so a single passed test says very little about the next one. Conventional machine safeguarding was built for deterministic machines: certifiable behavior verified through sensors, emergency stops and standards compliance. A model that generates actions on the fly cannot be certified that way.

The company was founded by Dr. Ding Zhao, who directs Carnegie Mellon's Safe AI Lab and previously worked at Google DeepMind; Kyle Wong, Safeworld's CEO, who founded Pixlee and led Stanford's StartX accelerator; and Simo Rachidi, formerly a principal security and machine learning engineer at Salesforce Einstein.

The product works like a stress test. A customer feeds in its robot's control policy, and Safeworld runs it through thousands of simulated scenarios populated with realistic human models — including rare and hazardous situations that would be expensive or dangerous to stage with physical machines. The platform integrates with physics simulators such as Genesis and MuJoCo, and targets the AI policy rather than the hardware. The pitch is a third-party evaluation that buyers and regulators can trust more than a manufacturer's own in-house test suite.

One customer is named so far: Gritt Robotics, which builds machines that install solar panels at industrial-scale farms — an environment that combines uneven terrain, changing weather and human crews working beside heavy equipment. Safeworld says it is also running pilots with automotive manufacturers, warehouse automation providers and medical device makers, none of which it has named, so the commercial evidence rests on a thin base until those convert to paid contracts.

The founders frame the need in blunt terms. "As robotics moves from impressive demos to everyday deployment, safety becomes a prerequisite for adoption," Wong said. Zhao put it as two problems at once: "The safety challenge is a combination of really advanced generative AI probabilistic evals and the trust part, and you need both to deploy a robot."

What it means: as humanoid and mobile robots move out of fenced cells and into shared workspaces, someone has to answer the question traditional certification never faced — how do you underwrite a machine whose behavior is sampled rather than fixed? Simulation-based validation is the most plausible answer, but it has real limits: simulated humans are approximations, a robot that passes ten thousand virtual scenarios can still fail on the one behavior nobody modeled, and liability stays with the operator and manufacturer no matter what a test report says. Safeworld's value will ultimately depend on whether insurers and standards bodies accept its outputs as evidence.