OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints

3 AgiBot https://omnimanip.github.io "Press button with hammer" "Press green button" "Insert flower into vase" "Insert pen into holder" "Close drawer" "Recycle battery" "Pour tea into the cup" "Pick up cup on dish" "Open drawer" "Put the lid on the teapot" "Close lid of laptop" "Open bottle" Passive Active Pour tea into the cup. Figure 1 . We proposed OmniManip, an open-vocabulary manipulation method that bridges the gap between the high-level reasoning of vision-language models (VLM) and the low-level precision, featuring closed-loop capabilities in both planning and execution.