Few-shot imitation learning for dexterous manipulation

Dexterous manipulation represents a fundamental challenge in robotics, demanding simultaneous force regulation, spatial reasoning, and real-time environmental adaptation. Traditional grasping pose selection methods often struggle with the combinatorial complexity of object pose difference and variability. In this work, we present a hybrid imitation learning framework that synergistically combines one-shot demonstration encoding with few-shot enhancement to optimize grasping pose generation for dexterous grasping systems. Our approach leverages geometric primitives derived from single demonstrations, augmented by ICP-based pose refinement, to generalize policies across objects with diverse geometries and material properties. The framework incorporates hierarchical collision detection and contact-aware trajectory optimization to ensure safety and stability during dual-arm coordination. Experimental validation across various object categories (in- cluding deformable, articulated, and irregularly shaped items) demonstrates an 94.4% accuracy for symmetric rigid objects. The system's adaptive capabilities are further evidenced by successful transfer to real-world tasks requiring centimeter precision. This work establishes few-shot imitation learning as a novel paradigm for dexterous object manipulation, bridging the gap between data-efficient learning and the physical constraints of real-world robotic systems. Our results highlight the critical role of geometric priors in few-shot policy adaptation.