Safeworld announced that it has emerged from stealth after raising a seed funding round exceeding $12 million, led by Shine Capital and a16z Speedrun, with participation from Box Group, Carnegie Mellon University Endowment, Innovation Endeavors and SV Angel. The company is working to build tools for assessing the safety of robotic control systems that rely on generative AI models.
Safeworld is led by Dr. Ding Zhao, director of the Safe AI Lab at Carnegie Mellon University, along with Kyle Wong and Simo Rachidi. The company starts from a fundamental problem in transferring generative models to robots: their behavior is probabilistic and less predictable than that of traditional algorithms, making them more complex to test before they operate near humans.
Simulating situations that are difficult to test in practice
Safeworld’s assessment involves creating realistic digital environments and then operating a simulated version of the robot using its actual control software. Thousands of scenarios are then conducted in which human models interact with the robot, including situations involving tripping and falling, or a person carrying boxes appearing near a blind corner in a factory.
These environments can be implemented using simulation models such as Genesis or MuJoCo. The aim is to measure practical factors, such as appropriate speed, stopping distance and the system’s ability to detect a person in an unusual posture, rather than merely demonstrating a robot operating under ideal testing conditions.
What changes in practice?
Robots differ from autonomous vehicles in that they operate in unstructured environments, and safety standards may vary from one facility to another. The people around them also do not behave uniformly; they may be standing, kneeling, running or falling, with differences in clothing, sizes, shapes and skin tones.
For this reason, Safeworld’s founders believe that testing a robot in edge cases cannot be reduced to a simple mathematical proof. Vishal Dugar, chief technology officer at Gritt Robotics, says that verifying the safety of systems must be done experimentally, especially when robotic arms operate alongside workers at solar-panel installation sites. Gritt Robotics has partnered with Safeworld to develop safety simulations.
Why does this development matter?
The problem the company is targeting lies at the intersection of two aspects: assessing the probabilistic risks of an AI system and building sufficient trust to enable robot deployment. Investors and founders believe that establishing safety practices before robots become widespread in homes and factories could help avoid discovering flaws after actual accidents occur.
Safeworld is also betting that robotics companies may need an external party to verify safety tests, and perhaps to share information about hazardous situations among competing companies. However, the company has not yet decided whether it will offer its platform as an external software service or adopt a services-based model, and it remains at an early stage of product development.
Simulation remains a testing tool, not an absolute guarantee of safety. The effectiveness of the results will depend on how realistic the human models, environments and scenarios being tested are, as well as on the tests’ ability to represent actual operating conditions when robots are deployed widely.