Dynamic grasping via multi-frame temporal association and transformer-based refinement

Abstract

Dynamic object grasping requires a robot to complete alignment, contact establishment, gripper closure, and lifting while the target pose evolves continuously under motion, occlusion, and sensing noise. In such settings, a key difficulty is not only fast control but also the temporal instability of grasp candidates: local grasp detectors frequently output candidate sets that jitter in position and orientation or switch identities across adjacent time steps, which prevents the downstream executor from maintaining a coherent grasping intention. This paper addresses that problem by introducing a temporal enhancement layer that performs multi-frame candidate association and lightweight motion refinement before grasp execution. The proposed framework first links grasp candidates across consecutive frames and maintains a compact set of stable candidate tracks, and then applies a lightweight Transformer-based refinement module to correct short-horizon motion deviations. To validate the resulting candidates under a deterministic and interpretable protocol, a candidate-guided staged executor is further constructed, consisting of pre-grasp alignment, short-range insertion, gripper closure, and lift-based verification. Closed-loop experiments are conducted in both static and dynamic settings, and the method is evaluated using task-level success rate, time-to-success, and sequence-level stability measures. The results show that temporally stabilized grasp candidates improve dynamic grasping robustness, reduce candidate jitter, and preserve a more coherent grasping intention in the presence of target motion.