andyzeng / visual-pushing-grasping

Train robotic agents to learn to plan pushing and grasping actions for manipulation with deep reinforcement learning.
http://vpg.cs.princeton.edu/
BSD 2-Clause "Simplified" License
898 stars 315 forks source link

TABLE 1 results #33

Open mike112223 opened 5 years ago

mike112223 commented 5 years ago

Hi, The first experiment is in a simulation where 30 objects are randomly dropped onto a table. From my experiences, 30 objects would make the workspace so crowded and I found that it was easy for some object to get out of the workspace at objects initialization. Therefore, that object could never be grasped and the mission couldn't be finished, which led to a low Completion. But the Completion in paper of VPG is 100%. Is there any trick to prevent objects from getting out of workspace, which you do not put into code? Or when do testing we should change something?

cjbarb7 commented 5 years ago

For the evaluation portion, you can provide a text file of the object types and which position within the workspace and orientation. There it would make it easier such that they are not outside the workspace. In the 'Creating your own test cases in simulation' part of the github it provides details on how to design such scenarios.

mike112223 commented 5 years ago

I know that part, but I just feel confused about the first experiment in paper, where 30 objects are randomly dropped onto a table, not deliberately designed. And the Completion reaches 100%.

cjbarb7 commented 5 years ago

In robots.py in the add_objects function, you could consider modifying it to randomly drop within the workspace but within a tighter boundary so it could prevent it from falling outside.

mike112223 commented 5 years ago

Actually, I just wanted to follow the paper and the original code to reproduce the results. As you say, I can modify the code, but then the code become not original. My point is that I just wonder how did the author achieve this 100%. Did he change the testing code or this result were not produced by this code?

cww97 commented 5 years ago

from the simulation we can see one thing like a boundary of a box out of the workspace which can make the blocks from going outside. We can move it to the worksapce. but I still cannot get 100%. and I dont konw where to compute the completion. From evaluate.py I can only get a 'clearance'