Real Life Implementation on Franka Emika Panda about panda-gym HOT 3 CLOSED

This sounds interesting,
Honestly, I have no experience with sim2real transfer, I just know that it is not straightforward.

First point, I understand that you want to work with velocity control. panda-gym works with velocity control: actions are velocity targets. So you don't need to modify the code of panda-gym.

Secondly, you should configure the interface with gym and ROS to use the learned model on the real robot. I wrote a code in the past that may help you: https://github.com/qgallouedec/franka_python . It is an interface between the real panda robot and python via ROS.
It is provided as is, I no longer maintain this repository.

If you find/write a clean and nice code that allows to control a real robot based on panda-gym, let us know!

from panda-gym.

Thank you for the response. I was looking into PyBullet documentation, and I guess there is this line of code here in the PyBullet.py ->
self.physics_client.setJointMotorControlArray(
self._bodies_idx[body],
jointIndices=joints,
controlMode=self.physics_client.POSITION_CONTROL,
targetPositions=target_angles,
forces=forces,
)

I think this needs to be changed to VELOCITY_CONTROL, please correct me if I am wrong. I did that and am trying to train the reach joint space sparse reach task, but there is no much improvement after 5.5 million timesteps.

from panda-gym.

Actually, it is indirect velocity control since the target position is current_position + action*dx.

See this line:

I don’t know if it is equivalent to velocity control in pybullet.

from panda-gym.