search

dalinel / ddpgBaxter

2 stars 1 forks source link

how did u reset the env while training?? #2

Open cc3941329 opened 5 years ago

cc3941329 commented 5 years ago

i'm using your code to finish a grasp job. I just want to know during training how can i reset the env in gazebo when the robot touches the block. besides, how can we define the situation of a successful grasp block?

dalinel commented 5 years ago

Hello, Did you try to execute both : roslaunch baxter_sim_examples baxter_ddpg.launch rosrun ddpg ddpg

baxter_ddpg.launch is using the gazebo_ddpg.py script in ddpgBaxter/src/baxter_simulator/baxter_sim_examples/scripts/gazebo_ddpg.py

You can see at line 320 of gazebo_ddpg.py that we reset the environment each time a message is received on the service /ddpg/reset

To successfully grasp a block you need to use a force sensor For more details about grasping you can also read the pdf file report.pdf, especially the section 4.2

cc3941329 commented 5 years ago

Hello, Did you try to execute both : roslaunch baxter_sim_examples baxter_ddpg.launch rosrun ddpg ddpg

baxter_ddpg.launch is using the gazebo_ddpg.py script in ddpgBaxter/src/baxter_simulator/baxter_sim_examples/scripts/gazebo_ddpg.py

You can see at line 320 of gazebo_ddpg.py that we reset the environment each time a message is received on the service /ddpg/reset

To successfully grasp a block you need to use a force sensor For more details about grasping you can also read the pdf file report.pdf, especially the section 4.2

that's so nice to get a reply from u. i finished the bug that i mentioned, however, i think that the robot get to the point with a very distort pose. i tried to add the angles as a punishment to the reward function, but it failed to converge. May i ask u that if u have any good ideas for that? what i am going to do is to add human demostration. but it seems hard.