"Bridging Imitation Learning and Compliant Manipulation: A Hier-archical Force-gated Framework for Robust Interaction"

"Constrained by stable, high-quality expert demonstrations and the maintenance of the robust action sequence of interaction dynamics, imitation learning for contact-rich tasks still faces significant challenges. To bridge this gap, a comprehensive imitation learning framework for manipulation tasks is introduced in this paper. First, an heterogeneous teleoperation system is developed to ergonomically and effectively acquire reliable expert demonstrations enriched with multimodal interaction force data. Then, to mitigate the temporal sparsity and phase-dependency inherent in Force\/Torque (F\/T) perception, a gating network is designed for the dynamic integration of F\/T modality into the action prediction Transformer. Furthermore, the predicted action sequences are coupled with a variable admittance controller in Cartesian space, dynamically regulating a stable and continuous physical interface during contact-rich operations. Finally, peeling experiments across diverse stiffness and target force validate that the force-gating mechanism significantly enhances success rates, while the hierarchical control with variable admittance substantially improves contact force stability. "