The robotics industry is shifting toward generative AI models for control, but this shift introduces significant safety risks due to the unpredictable nature of these architectures. Safeworld, a new startup, has been established to address this critical concern and ensure the safety of future humanoid robots.
Dr. Ding Zhao, director of Carnegie Mellon University’s Safe AI lab, co-founded Safeworld with veteran executive Kyle Wong and machine learning engineer Simo Rachidi. Their mission is to tackle the complexities of deploying safe, generative AI-driven systems.
“The safety challenge is a combination of advanced generative AI probabilistic evaluations and trust,” explains Zhao. “You need both to underwrite the risk of a probabilistic system and to deploy a robot safely.”
Safeworld is emerging from stealth with a seed round exceeding $12 million, led by Shine Capital and a16z Speedrun. Other investors include Box Group, Carnegie Mellon University Endowment, Innovation Endeavors, and SV Angel.
“The time to build an industry safety standard is now while robots are being designed and deployed,” advises a16z Speedrun partner Jonathan Lai. “By the time robots are in households colliding with children, it will be too late.”
Safeworld specializes in evaluating robotic control systems within high-fidelity simulations populated with realistic human models. While similar to the stress testing done by Tesla or Wayve for autonomous vehicles, this task is significantly more complex for robots operating in unstructured environments with varying safety standards.
Kyle Wong highlights common scenarios requiring rigorous testing, such as a blind corner in a factory. The company must determine the correct speed and stopping distance for a robot to avoid colliding with humans carrying boxes or moving unexpectedly.
To solve these problems, Safeworld builds digital versions of physical spaces in simulation engines like Genesis or MuJoCo. They integrate the robot’s real software and run thousands of scenarios to test how human models interact with the machine.
Simulating unpredictability is a major hurdle, as Zhao points out. “People are unpredictable,” he notes, citing tripping and falling as examples that are difficult to test empirically in the real world.
Although robot builders often possess similar tools, the founders believe there is a strong need for independent third-party validation. This external verification would allow competitors to share safety case data while ensuring the robot performs reliably at scale.
Vishal Dugar, CTO of Gritt Robotics, which builds AI for solar panel installation robots, agrees that formal mathematical proof is insufficient. His company is partnering with Safeworld to perform the necessary empirical safety simulations for their robots operating alongside humans.
Dugar emphasizes the complexity of human behavior that must be accounted for, such as kneeling, crouching, running, or tripping. The system must reliably detect humans in varying configurations regarding appearance, clothing, size, and skin color.
Safeworld is still in the early stages of development, determining whether to offer a platform for external users or a service-based approach. Despite this, the team is confident they are addressing a vital need in the industry.
“We’ll probably be the first profitable company in this field,” Zhao states. “Because if anyone wants to deploy, they need to pay us to handle the situation.”
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.