MODEL PREDICTIVE PATH INTEGRAL CONTROL FOR LEARNED DYNAMICS OF QUADCOPTERS

Teammates: Andrew and Ramya

In our final project for ESE 650, we implemented a model predictive path integral (MPPI) control algorithm that uses a neural network dynamics model for a CrazyFlie quadcopter to operate in a physics simulator. We trained the neural network to predict the quadcopter accelerations at each time sample in the simulated dataset of drone dynamics, before deploying the neural network as a dynamics model in the trajectory rollout step of the MPPI control algorithm. The controller is able to successfully complete navigation tasks and closely track desired trajectories. We compare our controller to a similar MPPI controller that uses an analytical dynamics model, which models the exact physics dynamics of the system. We demonstrate that our MPPI controller's performance approaches that of this controller. A direct future extension of this work is to demonstrate that for an environment with physical phenomena like downwash and drag, a neural network can encapsulate these complex, non-linear dynamics using a purely data-driven approach. This model would lead to increased controller performance over a naive analytical dynamics model.

For a detailed report and access to more videos, please see the writeup linked below. To check out the code base, click on the repository button.