How to generate my own training data?

[minolta001]

Hello. Nice work!

May I know how you generate training data for Maze2D?

[devinluo27]

Hello,

Thank you for your interest.

This folder contains scripts to generate training data.

At a high level, the dataset is a random walk in the environment -- we randomly sample a goal in the environment and use BIT* to generate a path from the current state to the goal state. This sampling procedure is repeated until the dataset reaches the specified length.

You can use this script for Maze2D. Note that this script utilizes CPU multiprocessing to increase speed, and you can set this num_cores according to your machine's configuration.

[minolta001]

Thanks for your response! This is very helpful.

May I ask you one extra question if you don't mind?

I can see your work also works with KUKA arm. In the code implementation, the outputs of your model are just the pose of end-effector and there is another lower-level controller to solve the joint control given end effector poses, or your work can train a forward kinematics model to solve joint space control?

[devinluo27]

For KUKA arm, our model's output is already a trajectory of the robot joint state (of shape [horizon, 7]), so I believe that lower-level controllers can be directly applied.

[minolta001]

Thank you! I guess then I should go through the KUKA code to see what's going on.

Again, thanks for your time!