by Spirit AI (Qianxun Intelligence)CN
Build one general robot brain that can learn new handling jobs from data rather than hand-written code, and sell it with Spirit's own robots.
Spirit v1.6 released June 2026; its RoboArena top spot was struck for benchmark gaming weeks later.
Updated 24 Jun 2026Checked 11 Oct0 updates this week
Camera frames and a task instruction are fed into a large neural network pretrained on images and text, so it already knows what objects and words mean.
The same network outputs short sequences of joint and gripper commands, repeated many times a second as the scene changes.
It learns from more than 200,000 hours of recorded real-world interaction data, then is tuned on specific jobs such as plugging in battery connectors.
| Spec | Spirit VLA models |
|---|---|
| RoboChallenge score (v1.5) | 66.09R (reported) |
| RoboChallenge task success (v1.5) | 50.33%R (reported) |
| Real interaction data (April 2026) | 200,000+ hoursR (reported) |
| Data target for 2026 | 1 million hoursR (reported) |
R reported by the company
Spirit AI (Qianxun Intelligence)
Spirit VLA robot models, Moz1 humanoid
Spirit AI (Qianxun Intelligence) is a Hangzhou start-up founded in 2024 that builds both robot brains and bodies: the Spirit series of vision-language-action models and Moz1, a wheeled humanoid working on CATL battery lines and in JD.com stores.
