Tight-clearance peg-in-hole for delicate components under non-rigid suction grasping via hierarchical imitation learning

High-precision automated assembly of industrial components with strict tolerances presents significant challenges due to the inherent instability of non-rigid suction grasping and the difficulty in acquiring expert demonstration data. To address these issues, an autonomous assembly framework is proposed, integrating specific data acquisition and force control strategies into a hierarchical imitation learning structure. First, Data acquisition efficiency is improved by recording simpler inverse pull-out tasks via Virtual Reality (VR) teleoperation and applying time-reversal. Second, the framework decomposes the assembly into “coarse approach” and “fine insertion” phases, utilizing the acquired heterogeneous datasets to train specialized visuomotor policies. Furthermore, a residual force policy, trained exclusively on limited real-world expert demonstrations with force feedback, is introduced to apply precise online corrections. This effectively suppresses the contact-induced tilting and grasp failure characteristic of non-rigid suction cups. Experimental results on real-world components with 100µm and 150µm clearances demonstrate an initial 85–90% success rate. Subsequently, an iterative refinement strategy aggregates deployment data to update the residual policy, further boosting the success rate to 90–95%. Ablation studies confirm the necessity of the framework components, validating the system's robustness.