Adapting Neural Robot Dynamics on the Fly for Predictive Control

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.

Figures (4)

• Figure A1: Quadrotor robot adapting to a payload equal to $35\%$ of its weight on the fly while tracking a reference trajectory.
• Figure D1: Overview of our approach for on-the-fly neural dynamics learning and predictive control.
• Figure D2: Range of values and their densities in the collected dataset of quadrotor positions, orientations, linear velocities, and angular velocities. The top row corresponds to the training set, and the bottom row to the validation set.
• Figure E1: Quadrotor tracking lemniscate and circular reference trajectories with an added $350\,\mathrm{g}$ payload ($35\%$ increase). (a–b) lemniscate without/with adaptation, (c–d) circle without/with adaptation. In (e-f), the transparent quadrotors denote the starting states while the bold ones denote the end states.