An AI that plays NFS Rivals

Note: "speedyblackman" is YT/Blackpanthaa's IGN. Nothing racist.

Here's the plan (Oct 13 2021):

• Three different models on three different datasets.

  1. Acceleration, break, reverse frames
  2. Left, right turn frames

• Work on left/right dataset first, then acceleration/break dataset.

• Implement speedyblackman.

Vague TODO:

• Create dataloader for the dataset working with and save the copy to drive.
• Test the dataset on different models.

Training completed (Nov 30 2021).

• Trained the turns data on ResNet18, VGG-19_bn, InceptionV3, MobileNetV2 and AlexNet.
  • AlexNet gave the best results when implemented.
  • Although the train, test accuracies were high, all the models seemed to overfit over a few epochs except AlexNet.
• Trained the acceleration data on ResNet18, MobileNetV2 and AlexNet.
  • MobileNetV2 gave the best results when implemented.
  • Almost all the models overfit. So the model that least overfit was chosen for the implementation.
• The data was not sufficient
  • 10GB (~52k frames) of turns data was used (which gave pretty good results).
  • 1.35GB (~6.8k frames) of acceleration data was used (which was definitely not sufficient and hence every model overfit).

Wanna play around with it?

1. Get NFS Rivals - you know what to do

2. Clone this repository and change directory into it

   git clone https://github.com/insaiyancvk/speedyblackman
   cd speedyblackman
   

3. Download the weights from drive and save them in speedyblackman folder

4. Install the dependencies

   pip install torch torchvision numpy keyboard opencv-python

5. Start the game and run the code

   python speedyblackman.py

Enough of my rant, look the result for yourself

Check it on youtube and few other clips in drive