Explainable Hypergraph Mamba Model for Mobile Manipulator Shelf Picking

Source abstract

Mobile manipulators operating in narrow shelf aisles must jointly reason about object affordances, shelf geometry, base placement, arm motion, and changing visual evidence. This paper proposes an explainable Hypergraph Mamba model in which each candidate picking event is represented by typed hyperedges linking shelf cells, target and obstacle items, grasp poses, base states, robot configurations, and short execution histories. Hypergraph aggregation captures higher-order physical constraints, while a selective state-space backbone retains action-relevant evidence with sequence complexity that scales linearly rather than quadratically with token length. The model also ranks influential hyperedges and temporal gates to provide decision-linked diagnostic rationales. Evaluation used 4,800 simulated shelf scenes and 620 physical robot trials covering five object categories, three shelf heights, and four clutter levels. Hypergraph Mamba achieved a 91.8% pick success rate, compared with 86.7% for the graph-transformer planner and 83.9% for the recurrent graph policy. Mean planning latency was 31.4 ms rather than 48.6 ms for the graph transformer, and the collision-contact rate was 3.6% rather than 5.2%. In failed-case audits, one of the three highest-ranked relations agreed with the engineer-assigned rationale in 78.6% of cases. These explanations indicate diagnostic consistency with modeled relations rather than complete physical causality. The results show that higher-order relational encoding and selective state retention can improve the efficiency, safety-related screening, and traceability of mobile-manipulator shelf picking under clutter and partial occlusion.