Explicit goal markers anchor zero-shot robotic placement

AnySlot treats language grounding as a separate visual target before control. The system edits the scene with a colored marker on the intended slot, lifts that point into 3D, and projects it across camera views. The action policy then follows this explicit goal. That design fits the task: slot placement needs both semantic choice and tight geometric alignment, and the paper argues that bundling both into one flat policy makes dense layouts hard to handle.

The benchmark pressure is also clear. SlotBench asks for slot selection under tight tolerances, with about 3 cm slots and correctness defined within 2 cm of the true slot center. It spans nine reasoning categories, including ordinal, negation, affordance, and world knowledge. On that setup, the paper reports nearly 90% average success, while a visible flat Diffusion Policy baseline reaches 16% on ordinal reasoning and 0% on the other shown categories. The available excerpt does not expose the full comparison table, so the strongest grounded claim is that the task sharply separates explicit-goal control from flat baselines in the shown results.