We tried to ameliorate the problem of manual setting of Artificial Potential Field constants in the APF woven implementation of the classic RRT* algorithm ( https://github.com/Snehal-Reddy/RRT_Star_With_APF ).
We generated 1,00,000 images with random obstacles.They can be accessed at this link- https://drive.google.com/open?id=1_RFMAEEHSNluhC_CLTgPwLQ8VW1Pq-zF
We then used BFS and Dijkstra to calculate actual shortest path (since it was unweighted).
Using the above data and using the area of the obstacle and the centre of the obstacle as features , we designed the cost function by comparing the predicted path after an epoch and the ideal path by comparing the path coordinate lists using mean square error. With this custom error function we built a neural network with the help of Keras. We had two hidden layers. We used 5 fold cross validation to improve the neural network. We also made use Adam's optimizer. Using this we predicted the potential function.
The results turned out pretty well and the tree converged well towards the target .
We use OMPL's inbuilt RRT-X to get the ideal path. We then build a neural network which learns by comparing its path with the RRT-X one without actually reverse engineering ideal potential constants from the ideal path. Research Paper is currently under development.