A Skill Transfer Method of Ultra-Micro Dispensing for Micro-Nano Manipulation Robots

To address the wide part variety, confined spaces, and complex positions in micro-assembly dispensing that hinder rapid adaptation to new parts, this paper proposes a dispensing needle approaching-alignment skill transfer method integrating imitation learning and meta-learning. The method first constructs a neighborhood sampling module based on a multi-head sparse attention mechanism and a position bias network, which extracts environmental states, and employs a multilayer perceptron (MLP) to predict the motion direction; the base skill model is trained with imitation learning and contrastive regularization losses. On this basis, a meta-learning framework based on the Reptile algorithm is introduced to train task-sensitive initialization parameters on simulated tasks, enabling adaptation to a new part with only 3–5 demonstration trajectories and limited gradient fine-tuning. To achieve simulation-to-real transfer, a U-Net performs three-class semantic segmentation of microscopic images; needle tip localization with a mean deviation of 1.8 px is achieved through principal component analysis (PCA) fitting; and the segmented map is fed into the adapted model to generate approaching trajectories in the real image space. Experiments on real microscopic images of eight parts show an average trajectory-generation success rate of 93.75%, average response time of 79.76 ms, average trajectory angular deviation of 2.48°, and terminal position deviation of about 2.60 pixels, with stable performance across different parts. The method effectively learns smooth, collision-free needle-tip approaching policies and generalizes efficiently to new parts, offering a feasible pathway for flexible micro-assembly dispensing.