GRU Model - Navigating

Описание: Playlist link:    • Deepbot Demos  

This playlist contains short clips of an e-puck robot in a simulated environment using the Webots open-source robotics simulator.

The robot is controlled by one of two neural network models (Conv or GRU model), or both (Combined model).

The Convolutional model is trained to turn the robot so as to find and lock on a target. It uses the robot's camera as input with 32x32x3 pixels.

The GRU model is trained to move the robot around and avoid obstacles. It uses the robot's 8 distance sensors.

Both models were trained using the Deep Deterministic Policy Gradient reinforcement learning algorithm.

The models are combined in the last two videos to chase a default e-puck that uses a simple example algorithm for obstacle avoidance, contained in Webots.

The project was conducted for the Deep Learning course of the Master's Degree in A.I. in Aristotle University of Thessaloniki, by Emmanouil Kirtas and Konstantinos Tsampazis with the help of Associate Professor at the Department of Informatics Anastasios Tefas.

Links:

e-puck: https://en.wikipedia.org/wiki/E-puck_...
Webots: https://cyberbotics.com/ , https://en.wikipedia.org/wiki/Webots
DDPG: https://arxiv.org/abs/1509.02971
Convolutional Neural Networks: https://en.wikipedia.org/wiki/Convolu... , http://yann.lecun.com/exdb/publis/pdf...
Gated Recurrent Units (GRUs): https://arxiv.org/abs/1406.1078

MSc: https://ai.csd.auth.gr/en/
Anastasios Tefas: http://www.aiia.csd.auth.gr/LAB_PEOPL...
Emmanouil Kirtas: https://github.com/ManosMagnus
Konstantinos Tsampazis: https://github.com/tsampazk