Step length and horizon?

[ludvikka]

Hello Baris! Great work on your master thesis! I am doing similar work for my master thesis and i am using some of your work as inspiration! We are using mujoco and Robosuite, and I am therefore not familiar with pybullet.

I have looked thorugh your code and thesis, but i am not able to see how long your steps are (the amount of virtual time that passes by in your simulation between the steps) and how many steps one epiode lasts?

It would be very helpfule if you are able to ansvew my questions.

Best regards Ludvik

[BarisYazici]

Hi Ludvika,

One episode lasts 150 timesteps. And one timesteps is ran over a duration of either 48 times or 24 times. When we reset the robot model I run the model 24 times positioning is also 24 times. Only open/close gripper actions are 48times.

This is where we define the timestep: https://github.com/BarisYazici/deep-rl-grasping/blob/59f605a44788c4c0e64bbe9ab1100e7a6b8ecc3b/manipulation_main/simulation/simulation.py#L38

Where we call the simulation helper: https://github.com/BarisYazici/deep-rl-grasping/blob/59f605a44788c4c0e64bbe9ab1100e7a6b8ecc3b/manipulation_main/simulation/simulation.py#L50

Where we call the run method from robot object: https://github.com/BarisYazici/deep-rl-grasping/blob/59f605a44788c4c0e64bbe9ab1100e7a6b8ecc3b/manipulation_main/gripperEnv/robot.py#L233

[ludvikka]

Thank you for your reply! I dont seem to fully grasp (pun intended) what you are saying. If one episode lasts 150 timesteps, and one timestep lasts 1/240 (sec*), then each epiosde lasts under 1 second? There is something I'm not understanding about this.

When you say that one timestep is ran over a duration of 24 times. To you mean that one timestep lasts 24*(1/240) seconds? Do you differentiate between timesteps and the self._time_step?