Cross-Scale SparseFormer-V2 Prediction for Dual-Arm Cable-Routing Operations

Source abstract

Dual-arm cable routing requires a predictor that connects global slack transport with local bending, friction, and fixture contact while operating within a closed-loop control cycle. This paper presents SparseFormer-V2, a cross-scale sparse model that receives synchronized RGB-D observations, cable-centerline samples, dual-gripper states, route waypoints, and candidate contact regions. Cable-centered tokens are organized at fine local, route-contact, and global coordination scales. Mechanically anchored sparse attention links these scales, temporal fusion retains short contact transitions, and shared prediction heads estimate future cable geometry, endpoint error, bend risk, contact state, and near-future routing feasibility over a 1.2 s horizon. Evaluation uses 18,400 simulated sequences and 2,160 real-robot sequences spanning six routing layouts, three cable diameters, two surface materials, and route clearances from 8 to 28 mm. SparseFormer-V2 achieved 92.8% routing-state accuracy, a 5.1 mm endpoint error, a contact F1 score of 0.887, and a success-forecasting F1 score of 0.903 at 18.7 ms per frame. Relative to a dense transformer, it reduced endpoint error by 40.7% and inference latency by 50.7%. Under 45% cable-span occlusion, accuracy remained 86.7%, compared with 77.9% for dense attention. Ablation results attribute the largest gains to cable-centered tokenization and cross-scale interaction. The results support SparseFormer-V2 as a real-time supervisory predictor for constrained cable routing, while self-contact, deep fixture cavities, and branched harnesses remain outside the demonstrated operating range.