Safeworld to Test the Safety of AI-Powered Robots

    Safeworld has announced that it has begun operations to test whether robots powered by generative artificial intelligence could harm people. The behavior of such robots cannot be predicted as accurately as that of traditional algorithms. The company will assess risks in simulations that recreate the behavior of real people. The goal is to identify dangerous situations before the robots are deployed on a large scale.    Safeworld was founded by Ding Zhao, experienced startup executive Kyle Wong, and machine learning engineer Simo Rachidi. Ding Zhao also leads the Safe AI lab at Carnegie Mellon University. The startup has raised more than $12 million (more than AZN 20.4 million) in early-stage funding. The funding was led by Shine Capital and a16z Speedrun. Another 5 organizations also invested in the startup.    The company tests robots’ control systems in a digital environment that recreates real human behavior. For example, a digital replica of a blind corner in a factory is created in Genesis or MuJoCo. The robot model, together with the software it runs, is placed in that environment. Thousands of different scenarios are then run to test situations in which it could encounter people.    The tests examine situations such as people falling, running, bending down, or stepping in front of the robot while carrying a box. Repeatedly testing such situations under real-world conditions is both difficult and dangerous. Gritt Robotics is working with Safeworld to develop safety simulations. The company’s robots help install panels at large solar power plants and work alongside employees.    Safeworld believes independent testing will be necessary in addition to manufacturers’ internal testing. This approach could make it easier for competing companies to share information about dangerous situations. The startup has not yet decided whether to offer its product as a standalone platform or as a service. The company is focusing not on prototypes prepared for demonstrations, but on robots that will operate at scale around inexperienced people.