Qianxun Intelligence Showcases Long-Horizon Robotics

机器之心 · wechat · 2026-07-17

This article highlights Qianxun Intelligence's embodied robotics capabilities showcased at WAIC, focusing not on single grasps but on the continuous execution of long-horizon tasks like "tidying up a living room".

In the live demo, Moz1 understands commands, breaks down tasks, and autonomously navigates to organize, clean, and adapt to sudden changes. For example, while processing a plushie, if a paper ball is tossed onto the table, it re-evaluates the environment, adjusts its sequence to throw the paper ball away first, and then resumes the unfinished organizing task. The article emphasizes that the difficulty lies in open, dynamic, unstructured environments where robots must not only grasp but also maintain task context, recognize anomalies, re-plan, and proceed after action failures or environmental shifts.

Regarding the tech stack and product deployment, the article notes that Spiritv1.6 adopts an integrated VLA and world model architecture and introduces abstract action tokens during pre-training. Meanwhile, Qianxun Intelligence accumulates real-world operational data through industrial scenarios, having validated non-standard processes like high-voltage test plug insertion on a CATL battery PACK production line with a stable success rate of over 99%. The Moz2, aimed at commercial service scenarios, is currently in the exploration phase, focusing on validating interaction methods, hardware form factors, and adaptability in public spaces.

Original post →

More from Companies & People

Companies & People channel →