Disclaimer: Project currently under development. Project done for the course "Introduction à la robotique mobile" (GLO-7021) of Université Laval, Québec.
In this project, neural dynamics Leaky-Integrate-and-Fire (LIF) and Adaptive-Leaky-Integrate-and-Fire (ALIF) from neuroscience are compared with a multilayer perceptron (MLP) from machine learning on a control task. The said task is to land a micro air vehicle (MAV) in a simplified environment designed with the Unity game engine. Artificial neural networks, all trained by reinforcement learning with the Proximal Policy Optimization (PPO) algorithm, show that the accomplishment of such a task is feasible, but fail to achieve results allowing to move to a more complex step. Indeed, it is shown that MLP succeeds in landing the MAV with a single degree of freedom using an optical flow camera and an event camera while the ALIF and LIF dynamics fail to do so the majority of the time. Finally, a development pipeline for accomplishing this goal is presented along with avenues of refinement for the continuation of the project.
The MAV used in this project is a quadrotor with a event camera attached to the front and a optical flow camera attached on the belly.
The environment contained a platform where the need to land.
In addition to the platform, some markers with random textures are placed randomly in the space.
Finally, the environment looks like this:
The event camera placed on the front of the MAV is used to detect the variation of intensity in the scene.
The neural dynamics used in this project are Leaky-Integrate-and-Fire (LIF) and Adaptive-Leaky-Integrate-and-Fire (ALIF). Those dynamics are implemented and come from the following project:
The pipeline for training the neural network is as follows:
The results of the training with one degree of freedom are as follows:
| Model | Inputs | Mean Rewards | Std Rewards |
|---|---|---|---|
| MLP-128x128 | P | 1.000 | 0.000 |
| MLP-128x128 | P+V | 1.000 | 0.000 |
| MLP-128x128 | OF | 0.000 | 0.000 |
| MLP-128x128 | OFH | 1.000 | 0.000 |
| MLP-128x128 | EventCam8x8 | 1.000 | 0.000 |
| MLP-128x128 | OFH+EventCam8x8 | -1.000 | 0.000 |
| LIF-10x5 | OFH | -0.080 | 0.997 |
| LIF-64x8 | EventCam8x8 | -0.980 | 0.199 |
| LIF-64x8 | OFH+EventCam8x8 | -0.940 | 0.341 |
| ALIF-10x5 | OFH | -0.300 | 0.954 |
| ALIF-64x8 | EventCam8x8 | -0.140 | 0.990 |
| ALIF-64x8 | OFH+EventCam8x8 | -0.960 | 0.280 |
- P: position of the MAV [x, y, z];
- V: linear velocity of the MAV [Vx, Vy, Vz];
- OF: optical flow represented in scalar;
- OFH: optical flow represented in one hot vector;
- EventCam8x8: event camera of shape 8x8.
pip install -r requirements.txt
From package manager:
- ML Agents Version 2.1.0-exp.1
From Assets Store
- OpenCvSharp+Unity
@article{Gince_MAVControlWithSNN_2022,
title={MAV Control With SNN},
author={Gince, Jérémie},
year={2022},
publisher={Université Laval},
url={https://github.com/JeremieGince/MAVControlWithSNN},
}