Can Safeworld convince people that gen AI robots won’t hurt them?

The growing trend in robotics involves entrusting robots to generative AI models, but this brings about a challenge: these advanced systems are not as predictable as traditional algorithms. Dr. Ding Zhao, leading the Safe AI lab at Carnegie Mellon University, has been tackling this issue for most of his career. Together with Kyle Wong, a veteran start-up executive, and Simo Rachidi, a machine learning engineer, Zhao established Safeworld, a company dedicated to addressing the safety concerns surrounding these generative AI robots.

Zhao explains that the safety challenge involves two main aspects: evaluating the probabilistic risk of the system and addressing the trust factor, both of which are crucial for successful deployment. Safeworld, emerging from stealth with a seed round of over $12 million, is focused on developing a platform to evaluate robotic control systems in simulations that include realistic human models.

This process is similar to what companies like Tesla or Wayve face when ensuring their vehicles respond appropriately to unexpected situations in real-world driving scenarios. However, robots operate in unstructured environments with varying safety standards across different locations. To assess the safety of a robot in a specific setting, Safeworld creates a digital replica of the environment using simulation tools like Genesis or MuJoCo, places the robot within the simulation, and then runs thousands of scenarios involving human interactions.

This task is particularly challenging due to the unpredictable nature of human behavior, such as tripping or falling. Safeworld's founders believe that third-party validation is necessary to share safety information between competitors, as robot builders need to consider all possible edge cases, including humans with different appearances, postures, and appearances.

Gritt Robotics, which develops AI systems for industrial applications, is partnering with Safeworld to validate their safety simulations. The company's robots work alongside human workers, and ensuring the robots do not harm them is a top priority. Safeworld is still determining the optimal approach for its product—whether it will be a platform for external users or a services-based model—but they are confident they are addressing the right problem, and they believe they may become the first profitable company in this field.