Learning Deep Nets for Gravitational Dynamics with Unknown Disturbance through Physical Knowledge Distillation: Initial Feasibility Study

TL;DR

This paper introduces a novel deep learning framework that combines prior analytical knowledge and observational data via knowledge distillation to efficiently model gravitational dynamics in high-DOF robots, even under unknown disturbances.

Contribution

It proposes a disturbance-adapting deep neural network framework using knowledge distillation that improves data efficiency and disturbance handling in robotic dynamic modeling.

Findings

Outperforms learning-from-scratch methods in data efficiency

Adapts effectively to unknown disturbances in high-DOF robots

Potential to surpass analytical models with increased training data

Abstract

Learning high-performance deep neural networks for dynamic modeling of high Degree-Of-Freedom (DOF) robots remains challenging due to the sampling complexity. Typical unknown system disturbance caused by unmodeled dynamics (such as internal compliance, cables) further exacerbates the problem. In this paper, a novel framework characterized by both high data efficiency and disturbance-adapting capability is proposed to address the problem of modeling gravitational dynamics using deep nets in feedforward gravity compensation control for high-DOF master manipulators with unknown disturbance. In particular, Feedforward Deep Neural Networks (FDNNs) are learned from both prior knowledge of an existing analytical model and observation of the robot system by Knowledge Distillation (KD). Through extensive experiments in high-DOF master manipulators with significant disturbance, we show that our method surpasses a standard Learning-from-Scratch (LfS) approach in terms of data efficiency and disturbance adaptation. Our initial feasibility study has demonstrated the potential of outperforming the analytical teacher model as the training data increases.