Event

SketchMate: VLM sketch model achieves zero-shot precision insertion

1 reports · 1 sources

Jiaming Wang and others propose SketchMate: letting a vision-language model build, for insertion tasks, a sketch model composed of a small number of geometric primitives and two mating coordinate frames, which is then executed by a standard planner and a compliant controller. The authors claim that in strict kinematic tests over 69 simulated insertion episodes, the method inserts 88% of parts at assembly poses, whereas ReKep and MOKA using the same VLM both failed to insert successfully. On new placements of eight tasks, the authors claim that reusing one stored run result once yields the same number of inserted parts as calling the agent per episode, and that generating the sketch takes 3.4 seconds rather than about 90 seconds. All the above results are simulation data reported by the paper's authors.