The research question addresses the challenge of building robotic manipulation systems by connecting perception, planning, and control through compatible representations and interfaces, especially when using VLM code generation.
The proposed approach, RIVET, integrates VLM-generated programs by leveraging a shared object-centric representation that includes per-object 6D poses and a relation graph, enabling the generation of cooperating perception, rendering, relation-inference, and planning programs.
The authors report evaluating RIVET on tasks like cube stacking, tangram rearrangement, and 3D assembly in both simulation and on a physical robot, achieving an 83% overall success rate with offline-generated systems.