TORCS AI Driver

TORCS controller of team "JesusTakeTheWheel".

Requirements

• TensorFlow
• keyboard (for optional manual data collection)

A requirements.txt for pip is provided.

Models

• Multi-Layer Perceptron with 3-layers with 150 hidden units each by default
• Recurrent Neural Network with 3-layers with 150 hidden units and sequences of length 5 each by default
• Extreme Learning Machine with a hidden layer of size 400

All models train using simple SGD or AdaGrad with a learning rate of 0.01.

Training drive data can be used to train the models using their respective training scripts:

$ python3 train/mlp.py input.csv output_path
$ python3 train/rnn.py input.csv output_path
$ python3 train/elm.py input.csv output_path

Optionally, the number of training parameters can be specified:

$ python3 train/mlp.py --iterations 200000 input.csv output_path
$ python3 train/rnn.py --iterations 200000 input.csv output_path
$ python3 train/elm.py --iterations 200000 input.csv output_path

The output folder will contain the fully trained TensorFlow model and can be loaded into the corresponding drivers.

Drivers

Jesus

The superclass for autonomous drivers. Driving commands are processed by the underlying "Holy Coast" ML-model.

• Default Drive behaviour uses commands from the currently engaged "Holy Coast" model
• Overtake Drive is used during overtaking and avoidance procesures of surrounding opponents
• Recovery Drive is used if the default controller is deemed to be stuck
• Cheesy Drive is engaged during opponent obstruction and "Cheesus" mode

Judas

Controller which drives down center track with a maximum speed of 80 km/h. While neither fast nor very intelligent, it collects very clean training data.

Manual

A manually controlled driver which collects sensor and command data for training autonomous models. Drive using WASD, press L to add data into save buffer and press J to clear buffer of recorded states. To record keypresses, this driver needs to be run as a superuser.

MLP, RNN, ELM

Autonomous driver using either an MLP, RNN or ELM as the underlying model.

Place any pre-trained models into the respective directories resources/models/mlp/, resources/models/rnn/ or resources/models/elm/. Setting Symlinks at these locations may help with the swift interchange of models.

Gear

The gear driver's special use case is the evolutionary optimization of up and downshift RPM parameters. Obtain these numbers by running a couple of generations and plug in the numbers into the Jesus driver.

Run

To start the default client simply run the start script:

$ ./start.sh

Optionally, a port the client should connect to can be specified (3001 is the default):

$ ./start.sh -p 3002

To run any specific driver, use the scipts located in run/ e.g.:

$ python3 run/mlp.py
$ python3 run/rnn.py
$ python3 run/elm.py

---

Christina, David and Max wish you successful autonomous racing!