Self-driving neural network car in GTA V – Charles 2.0

A self-driving car in GTA 5.

For more information on this project, and how it all began from simple lane detection to deep learning, follow the full tutorial series here:


  1. Even though this is 2 years ago, is it possible for your AI to use "C button" to look backwards, so your program can use the back view to train reversing, since real self-driving vehicles have front and back sensors.

  2. and make others ai and make a fake emotion that let them be one, this emotion control they about problem distances which isvery coscious

  3. can you make a neural network learn with what you make and consequences of somethomg

  4. Aren't you overfitting your validation set by tuning the final weights to match the class distribution perfectly?

  5. This neural network-based driver is very intriguing. Also, did you know that the AI gets special boosts when they drive fast to avoid failing scripted events that makes them very unfair. I found that pretty cool.

  6. why scale from 0-1? Remember that neural networks are non-linearized by activation functions. Making the distance of min and max values farther would be inefficient in modelling and prone to error and slower compared to a min-max distance of 1. Activation functions distort outputs making the error curve noisy. Using a data with a large margin of values (0-255) would also amplify the distortion of the error curve also amplifying inaccuracies in feature detection.

    This is also a must in using CNN's because of its kernelling and pooling method. Before the data pass through the classification neural nets it would undergo bunch of activation functions and filtering as well as pooling. By the time your features reach the classification neural nets, it would be noisy or has a great amount of loss. This loss could cause for a wrong node to activate and output node as well.

    Also normalization is not a harmful process. So the right question is, why not scale data from 0 to 1?

  7. I love the series you make and I LIKE HOW YOU'RE HONEST WITH YOUR VIEWERS. I wouldn't that honest with myself training a model, I'd cheat lol

  8. To help you improve:
    Decrease the angle theta by which the car turns when sees an object, theta should be more than diameter of the object.
    Speed control where speed=x, x depends on closeness of an object.
    Slowing down x(speed) on turns or angles which are higher than 50 degree and speed x is max at theta is=0 to theta = 10 degrees and object distance is equivalent of 50 m away(units as u wish)

    Activating horn where distance=d is less than equal to 10m

    I provided you with the physics and programming is on you and as a fellow coder these improvements will not take a lot of time but will make the ride a lot more smoother!

  9. you can augment data by flipping frames horizontally and flipping left/right button accordingly

  10. im telling you the real issue is the handling of the cars themselves being very poor, and the AI is doing its best. And it might not be able to see as good as it could if you used ReShade which allows you to hack into the shader itself and render whatever layer you want.

  11. You know, my only validation set is: If my car isn't on fire, I may keep driving it (not safely though – I (more often than not) drive worse than Charles).

  12. You could use your NN for vision and then use Model Predictive Control for path planning and the actual driving

  13. you can also teach your neural network by calling a taxi driver and learn from him ( it is an NPC that drive itself too)

  14. 😂😂God presenting his project about me for his higher dimension intelligence college assignment….

    God: 3:41

  15. Shouldn't the fitness just be based on 'always make the picture move'? If it's not moving, then he's stuck or dead. If it's moving, then he's moving–it's working optimally.

  16. Might just be a stupid suggestion. Relating to what you said about the tree problem, have you try feeding the model with solutions of getting out from different angles around the tree?

Leave a Reply

Your email address will not be published. Required fields are marked *