Deep Learning Helicopter Dynamics Models

Abstract

This paper pose the dynamics modeling problem as direct high-dimensional regression, and take inspiration from recent results in Deep Learning to represent the helicopter dynamics with a ReLU Network Model, a hierarchical neural network model.

This paper focuses on a method to learn a system dynamics model by leveraging captured state-action trajectories from expert demonstration.

Usually，the difficulties of system identification lie in three aspects:

1. stochasticity in the system
2. unstructured noise
3. the presence of unobserved state variables (the difficulty in modeling helicopter)

Modeling systems with hidden latent state is often treated as a parameter learning problem in a graphical model.

Takens Theorem states that for a large class of nonlinear dynamical systems with d-dimensional state space, only 2d+1 previous measurements of a single system output are required to effectively reconstruct the system state.

The main contribution of this paper is proposing a novel ReLU Network Model, providing details about parameter initialization and optimization methods.

The Helicopter System

The task of system identification is to find a function $F$ which maps from state-control space to state-derivative:
$$\dot{s}=F(s,u,\theta )$$

The helicopter state $s$ is represented by a total of twelve degrees of freedom.
The control state u u