Truck Backer Upper

Before Training

After Training

Training Process

Updates

• There is no manual process in the notebook anymore.
  • Lessons are now created automatically instead of manually. All we need to do is to specify the number of lessons and the first lesson configuration (optional) and the last lesson configuration (optional). After this, lessons will be created automatically.
  • Previously, I was specifying the maximum number of steps (max_steps) the truck is allowed to take manually. And I was doing this for each lesson separately. Now, I use a fixed number of max_steps for all lessons and it seems like it is working well.
• Stopping condition checks are not being tracked by autodiff anymore.
• The notebook is more clean now.
  • The reset() function in the Truck class is updated to reduce redundant code.
  • A new update_state() function is created in the Truck class. During the training of the controller, the state of the truck is now updated with this function.
• I added visualizations of gradients and loss values to Wandb instead of tracking them in Tensorboard and in the notebook. They can be seen in the links below:
  • Emulator Training
  • Controller Training
• I visualized the computational graph for the controller. The visualization of computatinal graph can be seen in computational-graphs folder and the visualization of the training process can be found in figures folder.

Notes

• The truck is still not able to back up when it is initialized in horizontal and reverse positions unless we use the configurations below:
  • self.θ0 = min(self.θ0 - self.θ1, deg2rad(80)) + self.θ1
  • self.θ0 = max(self.θ0 - self.θ1, deg2rad(-80)) + self.θ1