Alibaba Publishes the Qwen-RobotManip Technical Report for Robotic Manipulation

Alibaba's Qwen team introduced Qwen-RobotManip, a vision-language-action foundation model built on Qwen-VL, in arXiv:2606.17846. The approach aligns representation, motion, and action, with about 38,100 hours of pretraining data drawn from open datasets and human videos. The team reported first place on RoboChallenge, a 20% relative improvement over π0.5, and gains over prior leaders across all out-of-distribution settings tested.

Visit original source