nakuramino / deep-motion-planner Goto Github PK

@ Pratik
[from brainstorm.txt]

• Neural Network architectures (Designing)
  • 128x128
  • CNNs? nn?
  • RNN (Pratik)

I think it would be best if you design a CNN that takes in an image as the first input, and then the curr/goal state later on at the fully-connected layers. see #1 for more detail :)

• Setting the testing infrastructure so we can get statistical results
  • [DONE] % of successful paths generated from random start/goal states (Nishant)
    • done for ShootingStarNet and ADoubleStarNet (Nishant)
  • [CANCELLED] cost
  • [CANCELLED] planning time
  • [CANCELLED] anything you want to come up with that shows good/promising results
• [DONE] path, cost comparison for AStar (Nishant)
• path, cost comparison for non-holonomic RRT (Pratik)

Things to test:
[DONE] Normal planner (A* ShootingStarNet) (Nishant)
[DONE] Bidirectional Planner (A* ShootingStarNet) (Nishant)
[DONE] Normal planner (A* DoubleShootingStar) (Nishant)
[DONE] Bidirectional Planner (A* DoubleShootingStar) (Nishant)
Nonholonomic planner (RRT) (Pratik)

• test_data test_maps for both plannners (Mino)
• clean up what we have so it makes more sense to people
• implement some of the algorithms talked about in the MPNet paper

Meet up on Wednesday to finish things and write poster

• finish by Tuesday night, write everything on wednesday
• write report
• make presentation
• make presentation video

Instead of overloading the chat, I thought I'll keep my thoughts here:

input should be (curr_y, curr_x, goal_y, goal_x, map_image) and output should be action.

One issue I foresee is how the data should be returned. curr_state and goal_state are [2x1] arrays, but map_image is an 128x128 image, which means it can't be passed into a cnn as straightforward input (same can be said about normal neural networks). My guess is to overcome this by simply using tuples as our input. We take in a tuple, and then break it up in the forward function of our neural network, and then pass them in at appropriate locations of our neural net (i.e, map gets passed in at the convolutional layers, while the curr/goal state only get passed in at the start of the fully connected layers).

This is just a guess/suggestion tho, so let me know if you have any other ideas!