A robot arm can learn a screw-driving routine for thousands of hours, then lose most of that skill when moved to another arm. Alibaba's Qwen team is aiming at exactly that cost center with three embodied AI models.
A shared action language
RobotManip compresses different robot actions into an 80-dimensional standard action vector and uses switches to decide which dimensions apply. Describing motion as relative displacement from the camera view gives different arms a shared language.
That is where the headline number comes from: cross-configuration transfer success rises from the previous best 7.5% to 23.9%. Other held-out tests also improve, including 91.4% on LIBERO-Plus, 69.4% on RoboTwin-C2R hard and 45.6% on EBench.
Three models, not one
RobotNav covers instruction following, target navigation, object search, tracking and autonomous driving. RobotWorld uses a frozen Qwen2.5-VL plus a 20B diffusion Transformer to imagine future frames before the robot acts.
The data work is the real story
The most practical part may be data cleaning. One subset reportedly had 81% of segments rejected after checks for action jumps, timing errors and outlier values. If the result transfers to real production lines, the value is not a smarter demo, but skills that can move across bodies.
Sources: MarkTechPost, TechNode, CocoLoop; checks covered the three Qwen robot models, 23.9% transfer success, benchmark figures, 38,100 hours of manipulation data and the open-source status of RobotManip and RobotNav.