Adapting Neural Robot Dynamics on the Fly for Predictive Control

TL;DR

This paper presents a method for rapid online adaptation of neural robot dynamics models, combining offline training with efficient online updates for improved predictive control in real robots.

Contribution

It introduces a novel incremental neural dynamics model with low-rank second-order parameter adaptation for fast online updates, enhancing robot control.

Findings

Achieved robust predictive tracking on a real quadrotor.

Enabled rapid online model updates without full retraining.

Demonstrated effectiveness in novel operational conditions.

Abstract

Accurate dynamics models are critical for the design of predictive controller for autonomous mobile robots. Physics-based models are often too simple to capture relevant real-world effects, while data-driven models are data-intensive and slow to train. We introduce an approach for fast adaptation of neural robot dynamic models that combines offline training with efficient online updates. Our approach learns an incremental neural dynamics model offline and performs low-rank second-order parameter adaptation online, enabling rapid updates without full retraining. We demonstrate the approach on a real quadrotor robot, achieving robust predictive tracking control in novel operational conditions.