Spirit v1.5 has emerged as a defining milestone in the evolution of embodied artificial intelligence. Spirit AI has officially open-sourced its self-developed Vision-Language-Action (VLA) foundation model, Spirit v1.5, just days after it secured the top position on a leading global embodied AI benchmark. The move signals both technical maturity and a strategic push to accelerate innovation through openness.
The timing is significant. One day before the open-source release, Spirit v1.5 claimed first place on the Table30 leaderboard of RoboChallenge, the world’s first large-scale benchmark platform built specifically for embodied intelligence.
The model achieved a composite score of 66.09 and a success rate of 50.33 percent, becoming the only model to cross the 50 percent threshold and overtaking the previously dominant Pi0.5 model.
RoboChallenge is jointly organized by globally recognized institutions including Dexmal, Hugging Face, and the Beijing Academy of Artificial Intelligence. Unlike simulation-heavy evaluations, RoboChallenge emphasizes real-world testing on physical robotic systems.
This focus makes its results particularly meaningful for assessing whether embodied AI models can generalize beyond controlled environments and perform reliably in real operational settings.
A core strength of RoboChallenge lies in its real-hardware evaluation framework. The platform supports continuous testing on mainstream robotic systems such as Franka, Arx5, and UR5 single-arm robots, as well as dual-arm systems like ALOHA. Multi-view RGB and depth sensing enable precise measurement of how models perceive, reason, and act in complex physical spaces.
The benchmark evaluates more than isolated skills. Tasks are designed to test time-dependent reasoning, multi-stage long-horizon execution, and transfer learning across different robot configurations.
Models are also assessed on their ability to adapt to entirely new tasks, making RoboChallenge a gold standard for measuring general-purpose embodied intelligence rather than narrow task performance.
Spirit v1.5’s success did not come from excelling in a single metric. Instead, it demonstrated consistent, system-level improvements across perception, reasoning, and action. This balanced performance is what allowed it to surpass Pi0.5 and establish a new performance ceiling on the Table30 leaderboard. The results suggest a level of robustness that goes beyond benchmark optimization.
What further sets Spirit v1.5 apart is its validation in live industrial environments. In late December 2025, a humanoid robot named “Moz,” powered by Spirit v1.5, was deployed at CATL’s Zhongzhou manufacturing base. The robot entered real production on a new energy battery PACK assembly line, moving embodied AI from the lab directly into high-stakes industrial operations.
According to official data, Moz achieved a plug-in success rate exceeding 99 percent on the PACK line. It delivered three times the efficiency of human operators while significantly reducing high-voltage safety risks. At the same time, production quality remained stable, demonstrating that Spirit v1.5 is not only intelligent but also reliable under continuous operational pressure.
This real-world deployment is critical. Many embodied AI models perform well in controlled demonstrations but struggle with consistency over long periods. Spirit v1.5’s sustained performance at CATL indicates that it has reached industrial-grade engineering readiness, a threshold that few embodied intelligence systems worldwide have crossed.
The decision to open-source Spirit v1.5 amplifies the importance of this achievement. By making the model publicly available, Spirit AI is enabling researchers, developers, and robotics companies to build upon a proven, high-performing foundation. This approach accelerates experimentation, encourages transparency, and helps establish shared progress across the embodied AI community.
More broadly, Spirit v1.5’s success represents a shift in the global AI landscape. Chinese embodied AI models are no longer viewed as alternative or experimental options. With both benchmark leadership and real industrial validation, Spirit v1.5 positions itself as a preferred solution on the global stage, competing directly with the best systems worldwide.
This milestone also marks the first time a Chinese embodied AI model has been validated simultaneously through rigorous benchmark performance and large-scale industrial deployment. That dual validation places Spirit v1.5 firmly in the top tier of embodied intelligence systems and sets a new reference point for future development.
As embodied AI continues to move toward real-world utility, achievements like Spirit v1.5 highlight what is now possible. Open, benchmark-proven, and industrially deployed models are redefining expectations for robotics and intelligent systems across manufacturing and beyond.
For more updates on cutting-edge AI breakthroughs, robotics milestones, and global research trends, visit ainewstoday.org and stay ahead of the future of artificial intelligence.
