GN-RRT

Training code is adapted from my repo Torch-MTS.

GN-RRT vs. RRT:

Instruction

a_star_gen_images.py

generate n images, and p (start, end) pairs for each image.

in total n*p image samples.

default image shape: 200*200
gen_dataset.py

generate train, validation and test datasets for given n and p.

specify num_grids_height and num_grids_width in this file.

default: 20*20 grids, each grid is of 10*10 shape

train.py

train the model.

python train.py -m gridgcn -n 500 -p 20 -g <your_gpu_id>

grid_rrt.py

run NaiveRRT or GridNeuralRRT.

see this file for details.

Results

n=20, p=10, 200 images, max_iter=400

----- Grid Neural RRT -----
Avg time: 0.04270470976829529
Avg iters: 146.445
Success rate: 0.92

----- Naive RRT -----
Avg time: 0.08481519579887391
Avg iters: 272.04
Success rate: 0.66

n=100, p=10, 1000 images, max_iter=400

----- Grid Neural RRT -----
Avg time: 0.03659790062904358
Avg iters: 141.838
Success rate: 0.894

----- Naive RRT -----
Avg time: 0.08226466464996338
Avg iters: 272.459
Success rate: 0.629

n=300, p=20, 6000 images, max_iter=400

----- Grid Neural RRT -----
Avg time: 0.0367194459438324
Avg iters: 138.48533333333333
Success rate: 0.9016666666666666

----- Naive RRT -----
Avg time: 0.07920803066094717
Avg iters: 264.891
Success rate: 0.6476666666666666

注意数据的坐标系

在array和tensor中左上角是 (0, 0), 向下是第一维 (height维) 正方向, 向右是第二维 (weight维) 正方向
在画图的时候左下角是 (0, 0), 第一维画成了横轴, 向右是第一维正方向
实际上存数据的时候是一样的，只不过画图的时候把原点定在了左下角，所以一张图存的 (height, width) 但是画出来是 (width, height)
在做数据集的时候容易混乱, 然而就按 (H, W) 来就行, 不需要什么坐标转换; 不要惦记着画图是怎么画的, 容易把自己搞晕

XDZhelheim/GN-RRT

GN-RRT

Instruction

Results