May I ask whether your training process is carried out in simulation or directly on the robot arm

[l-okkkkk-k]

May I ask whether your training process is carried out in simulation or directly on the robot arm

[baifanxxx]

Thanks for your question. We train the high-level policy (RL) in simulation and run our hierarchical policy on a real robot arm.

[l-okkkkk-k]

What simulation software do you use, gazebo or CoppeliaSim

[l-okkkkk-k]

Besides, I see that the coordinate information of object blocks is obtained in your code. I have a question, since the two-dimensional coordinate information of object blocks is known, can this rearrangement problem be converted into a two-dimensional path planning problem

[baifanxxx]

Instead of using an off-the-shelf robot simulation environment, we used a grid environment, which can be viewed from the paper. The difficulty of the rearrangement problem lies in how to plan the order of different trajectories (the order of different objects will cause path cost) rather than the trajectory of an object. So path planning is a low-level problem, and rearrangement is a more advanced problem. A detailed introduction can be found in the paper.

[l-okkkkk-k]

I would like to ask a few different classes in env.py in your library function,ENV,ENV_ablation_wo_base,ENV_ablation_wo_repetition,ENV_ablation_wo_multi, what's the difference between them