mahyaret / kuka_rl Goto Github PK

Thanks for your codes..

I was wondering where can I write a code to command the kuka to relocate the block after grasping. whatever I do still the kuka joints returns the block to the start location of episode.

for example after grasping I wanted the joints get to a target location with help of calculateInverseKinematics and setJointMotorControlArray.

Could you please help me?
best

I believe that line 277 has a bug:

returns_current = values[-1].detach()

should instead say

returns_current = rewards[-1].detach()

will you be kind enough to share the pre-trained model in this repo?

Thanks a lot for your code. This is a solid and interesting project.

<1>I have run through your code in the jupyter notebook and I found you use "SummaryWriter" to inspect some scalars in the Tensorboard. Also I have tried the following code to inspect the whole PPO graph with "SummaryWriter" but I failed.

"""
dummy_input = torch.rand(1,3,40,40).to(device)
with SummaryWriter(log_dir = 'kuka-Diverse-ppo',comment='ActorCritic') as w:
w.add_graph(policy_net,(dummy_input,))
"""

I am not very familiar with the summarywriter, so could you give me more tutorials on this issue?

<2>Could you also teach us how to generate the training gif for pybullet-kuka-grasping as shown in your repo?
I am so interested in it.
Thank you so much !