Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On identifying the non-linear dynamics of a hovercraft using an end-to-end deep learning approach

Roland Schwan, Nicolaj Schmid, Etienne Chassaing, Karim Samaha, Colin N. Jones

TL;DR

The paper tackles nonlinear system identification for a novel hovercraft platform used to play air hockey, addressing how to learn a physics-inspired model directly from trajectory data. It adopts an end-to-end gradient-based approach with batched learning and a Runge-Kutta discretization, comparing a first-principles formulation to a learnable configuration-matrix representation. A key contribution is the monotonicity-regularized thrust model that enforces physically plausible behavior, enabling accurate short-horizon prediction and reliable reference tracking on real hardware. The work demonstrates a practical, data-driven framework for learning and controlling nonlinear dynamics with actuator delays, and releases datasets and code as open source to support broader adoption and extension.

Abstract

We present the identification of the non-linear dynamics of a novel hovercraft design, employing end-to-end deep learning techniques. Our experimental setup consists of a hovercraft propelled by racing drone propellers mounted on a lightweight foam base, allowing it to float and be controlled freely on an air hockey table. We learn parametrized physics-inspired non-linear models directly from data trajectories, leveraging gradient-based optimization techniques prevalent in machine learning research. The chosen model structure allows us to control the position of the hovercraft precisely on the air hockey table. We then analyze the prediction performance and demonstrate the closed-loop control performance on the real system.

On identifying the non-linear dynamics of a hovercraft using an end-to-end deep learning approach

TL;DR

The paper tackles nonlinear system identification for a novel hovercraft platform used to play air hockey, addressing how to learn a physics-inspired model directly from trajectory data. It adopts an end-to-end gradient-based approach with batched learning and a Runge-Kutta discretization, comparing a first-principles formulation to a learnable configuration-matrix representation. A key contribution is the monotonicity-regularized thrust model that enforces physically plausible behavior, enabling accurate short-horizon prediction and reliable reference tracking on real hardware. The work demonstrates a practical, data-driven framework for learning and controlling nonlinear dynamics with actuator delays, and releases datasets and code as open source to support broader adoption and extension.

Abstract

We present the identification of the non-linear dynamics of a novel hovercraft design, employing end-to-end deep learning techniques. Our experimental setup consists of a hovercraft propelled by racing drone propellers mounted on a lightweight foam base, allowing it to float and be controlled freely on an air hockey table. We learn parametrized physics-inspired non-linear models directly from data trajectories, leveraging gradient-based optimization techniques prevalent in machine learning research. The chosen model structure allows us to control the position of the hovercraft precisely on the air hockey table. We then analyze the prediction performance and demonstrate the closed-loop control performance on the real system.
Paper Structure (19 sections, 26 equations, 4 figures, 2 tables)

This paper contains 19 sections, 26 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Hardware Design
  3. Model
  4. Corrections due to shifted center of mass
  5. Thrust model
  6. Representation through a configuration matrix
  7. Control
  8. Learning from data
  9. Minimizing batched prediction loss
  10. Encoder design
  11. Regularization of force model
  12. Experimental setup
  13. Data preprocessing
  14. Models
  15. Learning
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Hardware overview of the hovercraft.
  • Figure 2: Schematic overview of the entire system.
  • Figure 3: Schematic overview of learning architecture.
  • Figure 4: Trajectories of reference tracking controller for position and orientation for models M0 (dashed) and M4 (solid) for three set point changes (top), with its corresponding control inputs for model M4 (bottom). The set points are indicated by the dot-dashed lines.