A Novel Robust Imitation Learning Framework for Complex Skills With Limited Demonstrations

Abstract

Imitation learning allows us to directly encode manipulation skills based on human demonstrations, facilitating rapid transfer of skills without any expert knowledge. Autonomous dynamic systems (DS) offer reliable stability and time-independence though sacrificing part of accuracy, and are increasingly attractive as an encoding method. As unstructured environments become more challenging, skill trajectories become more complex, and various disturbances are encountered, existing state-of-the-art encoding methods struggle to adapt to these complex tasks. This paper introduces a novel robust DS-based framework for learning skills in complex tasks, which consists of trajectory regularization, adaptive segmentation, skill modeling, and skill organization based on new task requirements. It achieves a task-level generalization so that the operator only needs to focus on the semantic deconstruction of a task. Additionally, we propose an online modulation policy for the skill decision engine to address two types of disturbances: enhancing convergence speed for large-scale disturbances and improving fitting capability for small-scale disturbances while still keeping stability. To evaluate the effectiveness of the proposed framework, we conduct various comparison experiments in simulation and a real-world sugar-scooping task to assess the generalization performance and the ability of resistance to disturbances. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —Imitation learning for complex tasks is crucial to the development of robot intelligence. However, achieving a balance between maintaining generalization accuracy and robustness remains a challenging problem that necessitates continuous exploration in the field of imitation learning. The purpose of this paper is to propose a robust imitation learning framework from human demonstration, which includes preprocessing, learning and generalization, that can be applied in industrial production or daily life. Considering appropriate trajectory segmentation and self-organization strategies can effectively improve the generalization accuracy by prior research, it is necessary to introduce them into our framework. Most importantly, we design novel disturbance-resistant online modulation strategies from both task-level and motion-level aspects. To validate the effectiveness of our approach, we conduct simulations and coffee scooping experiments. The results show that skills acquired through demonstration can reliably, accurately, and safely perform tasks even in uncertain environments. This paper is a systematic and pioneering attempt to implement.

Topics & Concepts