SafeWorld raises $12M to make humanoid robots safe ... before they meet real people
Robots are getting smarter, faster, and more autonomous. They're leaving the cages, moving into warehouses and hospitals and homes. And the people deploying them mostly have no systematic way to prove they're safe around people before something goes wrong.
SafeWorld wants to fix that.
The Carnegie Mellon spinout emerged from stealth this week with $12.2 million in seed funding and a platform for stress-testing AI-driven robots in simulation before they ever meet a real human.
The gap everyone is ignoring
Here's the uncomfortable reality: humanoid robots are shipping at scale, but safety validation isn't keeping up. Companies run some tests in-house, watch the demo go well, and ship. The edge cases, the weird lighting, the unexpected human movement, the failure mode at hour 47 of a 48-hour cycle: those get discovered in the field.
That's how you get incidents.
SafeWorld's pitch is that you should test those scenarios before deployment, not after. Its platform lets engineering teams build test scenarios in a browser, no simulation expertise required, then run millions of variations with reactive human motion models. The system tracks how a robot performs across all of them and generates a documented safety record.
Who's behind it
The founding team is about as credentialed as you'd want for a safety company.
- Ding Zhao, co-founder and director of CMU's Safe AI Lab, has spent 17 years on safe autonomous systems. He's a National Science Foundation Career Award winner and a former Google DeepMind researcher.
- Kyle Wong, CEO, previously built Pixlee, an AI and UGC platform used by over 1,000 brands that he took from inception through acquisition.
- The third co-founder is Simo Rachidi, an ML engineer.
Shine Capital and a16z Speedrun led the round, with Box Group, the Carnegie Mellon University Endowment, Innovation Endeavors, and SV Angel also participating.
CMU's endowment writing a check into a CMU spinout isn't unusual, but a16z Speedrun's participation is a strong signal about where the partner group sees the market going.
What the platform actually does
The core technical differentiator is reactive human motion simulation.
Most safety testing in robotics runs the robot against static or scripted scenarios: a human stands here, moves like this, robot responds. That's not how real deployment works. Real humans are unpredictable ... they trip, they flinch, they walk into a robot's path while looking at their phone.
SafeWorld's models simulate that variability at scale. Instead of testing against 50 scenarios, you test against 50,000 variants, helping you find the edge cases before they find you.
(At least, that's the theory.)
Early customer Anyware Robotics is already using it.
CEO Thomas Tang said the platform gives them "a scalable way to test challenging scenarios" and strengthen safety processes. That kind of language will resonate with every company trying to get past the liability question on enterprise deployments.
Why the timing is right
Consider the scale of what's coming. Toyota is planning 400,000 humanoid robots across its factories by 2032. Industrial robot installations globally just hit five million. And those are the controlled environments: structured factories, defined spaces, known hazards.
The harder deployments, homes, hospitals, construction sites, are still ahead. And for those, the fail-safe question isn't theoretical.
There's also a regulatory angle that's developing in real time. Singapore's $100M public safety robotics center is explicitly about training and validating robots for high-stakes missions. The EU AI Act has implications for autonomous systems in sensitive environments. As regulations tighten, companies that can produce documented safety evidence will have a significant edge over those that can't.
SafeWorld isn't just selling peace of mind. It's selling a paper trail that makes deployment defensible.
The deeper question
The hard version of this problem isn't hardware safety, it's software safety.
A bolt won't fall off because the model updated. But an AI-driven robot's behavior absolutely can change after a software update, a fine-tuning run, a new deployment environment. SafeWorld's continuous retesting capability, which reruns scenarios after each update, is the piece that addresses this.
Kyle Wong puts it plainly: "If you're going to have more capable robots," you need rigorous validation methods. That's the right framing. More capable is not the same as safer. The industry has treated them as synonymous for too long.
$12.2 million is a seed round, of course.
SafeWorld still has to prove the platform works at the complexity levels real deployments demand, convince risk-averse enterprise buyers to add a new vendor to their stack, and build fast enough to stay ahead of incumbents who will notice this market.
Those are not small tasks.
Sources
Written by
I tell stories with data and build startups. I've raised angel and VC funds, led software companies and startups, and chronicled the rise of the mobile economy. Currently I write for Forbes, host several podcasts, and am starting up HumanoidMegaHub. I was TUNE's mobile economist and Singular's VP of Insights, built the research team at VentureBeat, and managed teams creating software for partners like Intel and Disney. In past lives I built websites, e-learning platforms, and mobile apps, and, a long time ago, was named to Folio's top 100 of the media industry's "most innovative entrepreneurs and market shaker-uppers.”