search

AutoMecUA / AutoMec-AD

Autonomous RC car with the help of ROS Noetic and ML.
GNU General Public License v3.0
15 stars 2 forks source link

Autonomous Driving Reinforcement Learning #124

Closed Joaoprcf closed 1 year ago

Joaoprcf commented 2 years ago

We need a reward system given by gazebo services. The first Reinforcement learning technique applied will be ARS (Augmented Random Search), which will be improved with a 2-3 layer neural network after the initial training (step 2), this new structure will have its weights on the first layers and the last one can be retrained again with ARS. The inputs given to this model can be obtained through CNN preprocessing and others. A script with capabilities to spawn and delete models is required.

manuelgitgomes commented 2 years ago

Revise requirements.txt.

manuelgitgomes commented 2 years ago

ROTA implementation should also be studied @brnaguiar

manuelgitgomes commented 2 years ago

For the referee try researching for frenet coordinates: https://www.selfdrivingcars360.com/glossary/frenet-coordinates/

manuelgitgomes commented 2 years ago

Search for transformers use in autonomous driving

callmesora commented 2 years ago

Search for transformers use in autonomous driving

Vision transformers are kinda new, they have been conquering the Deep learning world, however, they are still not the best fit for real-time applications due to their high computational requirements. I doubt that we could take any advantage by using them, especially with our hardware.

manuelgitgomes commented 1 year ago

Study the code in: https://github.com/AutoMecUA/AutoMec-AD/blob/fec62272809c74442c5ee7295ec3230bd90147ed/prometheus_driving/scripts/ml_training.ipynb

This is a CNN and we are trying to change it to an RNN.

PedroMS3 commented 1 year ago

For the moment being there is no member capable of working on this issue, therefore if we could get any member that knows reinforcement learning would be great