Robot safety gets its evidence layer: SafeWorld exits stealth with $12M

SafeWorld emerged from stealth on October 5, 2026 with $12M in seed funding to build safety testing and simulation software for robots operating around people. Shine Capital and a16z Speedrun co-led the financing, with BoxGroup, Carnegie Mellon University Endowment, Innovation Endeavors, SV Angel, Ovo Fund, Valkyrie, Zelda Ventures, Alpha Square Group, Founders Future and Brave Capital participating. The Palo Alto company did not disclose a valuation, previous institutional financing, or a precise allocation of the new capital.

The company was founded in 2025 by Kyle Wong, Ding Zhao and Simo Rachidi, a team spanning startup operations, autonomous-system safety research, cybersecurity and production machine-learning infrastructure. Wong is co-founder and CEO; he previously founded Pixlee, an AI and user-generated-content platform acquired in 2022, and later served as CEO of StartX. Zhao directs Carnegie Mellon University's Safe AI Lab, where his research examines trustworthy AI and safety for physical human-robot interaction. Rachidi is co-founder and CTO.

The problem SafeWorld targets gets more expensive as robots get more capable. Traditional industrial robots often work inside controlled spaces with movements engineers can define in advance. A robot that perceives its surroundings and adapts its actions can work in warehouses, factories, hospitals and construction sites, but probabilistic behavior makes every software version, sensor configuration and workplace layout a new safety question.

The dangerous cases are also the least responsible to reproduce around people. A worker may step from behind a cart, kneel in an unexpected place, carry an object that blocks a sensor, or fall into a robot's path. Lighting, clothing, body shape, equipment placement and approach direction can all change what the robot perceives. Physical testing remains necessary, yet it is slow, expensive, and incapable of covering every combination on its own.

SafeWorld's technical approach begins with a defined task, hazard, requirement, scenario, measurement and acceptance rule. Its software generates controlled variations, places reactive human models inside simulated environments, runs the robot's control system through those cases, and preserves the configuration and results so failures are easier to discover, compare and revisit after software changes.

The company is explicit about the limits of that evidence. A simulation result is evidence about a defined test, not a universal declaration that a robot is safe; a perception model can detect a person while the larger system still reacts too late, and a digital human can reproduce an occlusion pattern while failing to represent every real posture, sensor artifact or physical response. SafeWorld says simulation does not replace physical validation where the application or safety standard requires it.