Source note
AnySlot: Goal-Conditioned Vision-Language-Action Policies for Zero-Shot Slot-Level Placement
Vision Language ActionGoal Conditioned PolicyZero Shot ManipulationSlot Level PlacementSpatial Reasoning
Summary
AnySlot targets slot-level robotic placement from natural language, where the robot must choose the exact slot and place with sub-centimeter precision in unseen layouts. It does this by converting the instruction into an explicit visual goal marker and then using a goal-conditioned VLA policy to execute the placement.
Problem
- The paper studies zero-shot slot-level placement: given a compositional language instruction, the robot must identify the correct slot among dense candidates and place the object accurately enough for physical success.
- This matters for tasks such as precision assembly and factory automation, where choosing the wrong compartment or missing alignment by a small margin causes failure.
- Existing flat VLA policies mix language reasoning and motor control in one model, while prior modular systems often reduce the target to a single coordinate, which loses slot shape and boundary information needed for precise execution.
Approach
- AnySlot uses a two-stage pipeline. A high-level grounding module takes the initial scene image and the language instruction, then edits the image by placing a colored sphere marker on the target slot.
- The system extracts the marker center, lifts it into 3D with the depth map and camera calibration, and re-projects it into all camera views as a consistent overlay. This overlay is the visual goal.
- A goal-conditioned VLA policy, built on π0.5 with a PaliGemma-3B backbone and a flow-matching action expert, receives the multi-view observations plus the overlaid marker and outputs continuous robot actions.
- The low-level instruction is fixed to a simple template such as moving the object into the slot with blue-sphere guidance. Language-based slot reasoning stays in the high-level module; the low-level policy focuses on execution.
- The paper also introduces SlotBench, a SAPIEN simulation benchmark with 9 task categories for slot-level reasoning: ordinal, size, height, distance, compositional relations, negation, vague language, affordance, and world knowledge.
Results
- The paper claims AnySlot achieves nearly 90% average success rate on SlotBench across the 9 zero-shot slot-placement task categories.
- SlotBench uses tight geometry: slot size is about 0.03 m, object length about 0.15 m, and instruction accuracy counts a prediction as correct only if it is within 0.02 m of the ground-truth slot center.
- The low-level policy is trained on a synthetic simulator dataset with object pose randomization up to 0.05 m translation noise, then evaluated zero-shot on unseen layouts and instructions.
- In the visible table excerpt, a flat Diffusion Policy baseline reaches only 16% success on ordinal reasoning and 0% on the other shown categories, which is consistent with the paper's claim that flat baselines fail on most SlotBench tasks.
- The excerpt does not include the full quantitative table for all baselines and AnySlot, so exact per-category gains, baseline averages, and comparison margins are not fully recoverable from the provided text.