From Numerals to Letters: A Meta Learning-based Robotic Calligraphy System

Robotic calligraphy is often regarded as an important automation benchmark problem due to its theoretical significance of integrating aesthetic creativity to accurate robotic task implementation, in addition to its practical impact in culture preservation and promotion. Despite promising results achieved through technological advances, an unsolved key challenge is the requirement of a large amount of training time and data sampling for various fonts. This paper presents a robotic meta learning approach, which is able to effectively transfer the writing ability of one type to that of another. The proposed meta learning system is realized using a meta learner and a meta discriminator. The meta learner generates robotic writing actions, and the meta discriminator provides feedback to the writing result based on loss information. The meta learner is implemented by combining Long-Short-Term Memory (LSTM) and generative adversarial networks, with the support of a multi-head attention mechanism to detect similar features amongst various tasks. The proposed approach implements an inner loop and an outer loop to update the parameters of the meta learner using the feedback provided by the meta discriminator. In particular, the inner loop optimizes the learning of single numeral writing task, and the outer loop fine-tunes the meta information for all tasks. In the meta test phase, the meta generating model learned from the numerical tasks is applied to writing letters. Experimental results prove that under the meta learning framework, the robot can use considerably less training time to learn to write various types of characters with the guidance of meta information, as compared to those methods of learning new fonts from scratch.