I'm using a model car which can be remotely controlled and which has a front facing camera.

My goal is to train a CNN which would be able to efficiently park this car in a simple environment (pictured below).

enter image description here

During the training process I would put the car in a random spot in my room (with the blue box visible on the camera) and I would park the car in the spot with the blue box (I'd switch the box every 10 or so parking "episodes"). While driving I save an image alongside the steering value ([-1,1]) and throttle value ([-1,1]) at a rate of 10 images (frames) per second. Training set consists of 10_000 images and the validation set consists of 1000 images.

Using this approach I tried training several different CNNs (mine, ResNet18, ResNet34, ResNet50...) but all the results were fairly disappointing. The only thing I can say is that the car learned that it should stop when it gets close enough to the blue box.

This leads me to believe there's a fault somewhere in my approach. I would be grateful if someone would be willing to share any tips or even better if someone has some experience with the goal I'm trying to achieve.

  • $\begingroup$ You could try a different approach. Your problem seems to be the same as the ones with autonomous cars. Why not use an embedded camera, real-time image recognition like Yolo, and a Reinforcement Learning process? $\endgroup$ Jun 7 at 6:25


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.