In January 2026, startup Spirit AI from the Chinese mainland rocketed to the top of the RoboChallenge real-world robotics benchmark with its latest foundation model, Spirit v1.5. 🤖 With a total score of 66.09 and a task success rate of 50.33%, it surpassed the pi0.5 model by U.S.-based Physical Intelligence to claim first place.
Spirit AI has open-sourced Spirit v1.5 and its resources, inviting developers to build on its unified Vision-Language-Action architecture that blends perception, reasoning, and action into one end-to-end system. This approach aims to reduce errors common in traditional modular designs.
RoboChallenge is often called the “global exam” for robots. It evaluates embodied intelligence models in real-world settings through 30 tasks—everything from object placement and target recognition to tool use. Spirit v1.5 was the only model to clear a 50% success rate, making this achievement even more impressive. 🎯
Spirit AI is based in Hangzhou, capital of Zhejiang Province on the Chinese mainland—also home to AI startups like DeepSeek and robotics firm Unitree Robotics. In June 2025, the company unveiled Moz1, a humanoid robot designed for logistics and industrial use.
Associate Professor Qiu Jiefan from Zhejiang University of Technology noted, “For embodied intelligence, the ability to understand and perform across multiple tasks and scenarios is very important,” adding that while large-scale deployment is still on the horizon, this result is a giant leap toward practical service robots.
Founder and CEO Han Fengtao envisions even more service robots joining industries within the next two to three years. With Spirit v1.5’s success, the future of robotics feels closer than ever. 🚀🔧
Reference(s):
Chinese AI startup tops global embodied intelligence benchmark
cgtn.com
