A neural network based framework for variable impedance skills learning from demonstrations

Abstract

Robots are becoming standard collaborators not only in factories, hospitals, and offices, but also in people’s homes, where they can play an important role in situations where a human cannot complete a task alone or needs the help of another person (i.e., collaborative tasks). Variable impedance control with contact forces is critical for robots to successfully perform such manipulation tasks, and robots should be equipped with adaptive capabilities because conditions vary significantly for different robotic tasks in dynamic environments. This can be achieved by learning human motion capabilities and variable impedance skills. In this paper, a neural-network-based framework for learning variable impedance skills is proposed. The proposed approach builds the full stiffness function with the acquired forces and position learned from demonstrations, and then is used together with the sensed data to achieve the variable impedance control. The proposed algorithm can adapt to unknown situations that change the learned motion skill as needed (e.g., adapt to intermediate via-points or avoid obstacles). The proposed framework consists of two parts: Learning motion features and learning impedance features. The motion features learning is validated by reproducing, generalizing, and adapting to transit points and avoiding obstacles in the LASA dataset. Impedance features learning is validated based on a virtual variable stiffness system that achieves higher accuracy (approximately 90%) compared to traditional methods in a manual dataset, and the whole framework is validated through a co-manipulation task between a person and the Franka Emika robot.