January 12, 2026 — Spirit AI has formally open-sourced its self-developed VLA (Imaginative and prescient-Language-Motion) basis mannequin, Spirit v1.5.
Simply someday prior, Spirit v1.5 claimed the highest spot on the Table30 leaderboard of RoboChallenge, a worldwide benchmark platform for embodied intelligence fashions. The mannequin achieved a composite rating of 66.09 and a hit price of fifty.33%, making it the one mannequin to surpass a 50% success price and surpassing the beforehand dominant Pi0.5 mannequin.
RoboChallenge is collectively organized by internationally acknowledged establishments together with Dexmal, Hugging Face, and the Beijing Academy of Synthetic Intelligence (BAAI). It’s the world’s first large-scale benchmark platform purpose-built for embodied intelligence robots, emphasizing real-world testing in bodily environments to carefully consider fashions’ sensible generalization capabilities.
A key energy of RoboChallenge lies in its real-hardware analysis framework. The platform helps 24/7 testing on mainstream robotic techniques, together with single-arm robots comparable to Franka, Arx5, and UR5, in addition to dual-arm techniques like ALOHA. Multi-view RGB and depth information permit exact evaluation of how VLA algorithms generalize in bodily environments. Take a look at eventualities cowl essential dimensions comparable to time-dependent reasoning, multi-stage long-horizon duties, and consider switch effectivity throughout completely different robotic configurations and fully new duties—making RoboChallenge a gold-standard benchmark for general-purpose robotic intelligence.
Notably, Spirit v1.5 didn’t surpass Pi0.5 with a single standout functionality, however by means of complete, system-level efficiency positive aspects. Humanoid robotic “Moz,” powered by Spirit v1.5, was deployed in late December 2025 at CATL’s Zhongzhou manufacturing base, the place it entered dwell operation on a brand new vitality battery PACK manufacturing line.
In accordance with official information, Moz has achieved a plug-in success price exceeding 99% on the PACK line, delivering thrice the effectivity of human operators whereas considerably decreasing high-voltage security dangers and making certain steady manufacturing high quality. The mannequin’s sustained and dependable efficiency at CATL demonstrates that Spirit v1.5 has reached industrial-grade engineering readiness.
Spirit v1.5’s high rating and real-world deployment characterize greater than a single-model breakthrough. They sign that Chinese language embodied AI fashions are transitioning from “various choices” to “most popular options.” Extra importantly, this achievement marks the primary time a Chinese language embodied mannequin has been validated each on measurable benchmark efficiency and in actual industrial eventualities, inserting it firmly within the world high tier of embodied intelligence techniques.
Supply: JizhiUltra
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the newest breakthroughs, get unique updates, and join with a worldwide community of future-focused thinkers.
Unlock tomorrow’s traits right this moment: learn extra, subscribe to our e-newsletter, and change into a part of the NextTech neighborhood at NextTech-news.com
