Release code for Mars Navigation and Mujoco continuous control ?
zuoxingdong opened this issue · 9 comments
Is it also ok to release the code for Mars Navigation dataset, models and particularly the one for continuous control via Mujoco simulator. Did you use OpenAI Gym for it by the way ?
In addition, there are other continuous control task in OpenAI Gym, e.g. LunarLander, CarRacing, BipedalWalker, do you think we can try to use VIN on these tasks ? Perhaps even for some games in OpenAI Universe ?
Is it possible for you to share which image you used from the HiRISE dataset? In the paper you mentioned you used one image with multiple crops of size 128x128 pixels. Since I am working on implementing this myself it would be helpful to know which image it was so as to try and mimic the results of the paper.
Hi Kent,
Here's the image + elevation data. I also include the MAtlab file for preprocessing it, and the script for generating an MDP out of it. It is not cleaned up though...
This is awesome, thank you so much!
Could you release the Mars Navigation dataset used in your paper? Thank you!
@avivt
Thank you for releasing the preprocessing code for Mars Navigation. Could you share the script for generating an MDP? I got into trouble when I tried to use the 'obstacle_mars' class. It is not clear how to use it to generate data for training VIN?
This is planned to happen eventually. The code is currently too messy (deadline code...) to directly upload, and I'm personally too busy at the moment to clean it up. But it's on my list! For the continuous task,I didn't use the Gym. I built on the GPS codebase, which also supports Mujoco agents. But I guess gym could be easily used as well, as VINs are only a NN architecture. Regarding other domains: I'm pretty sure that with a bit of thought on the design (e.g., planning state/action spaces) VIN can be applied to such domains. The main benefit of VINs though, as I see it, is generalizing to changes in the domain that were not seen during training. This was very clear in the navigation with obstacles, where a change of domain meant a change of obstacles. In the domains you mention, I'm not sure whether the domain can change at all. Perhaps the car racing with different tracks would be cool to show.
…
On Fri, Dec 23, 2016 at 3:26 PM, Xingdong Zuo @.***> wrote: Is it also ok to release the code for Mars Navigation dataset, models and particularly the one for continuous control via Mujoco simulator. Can we use OpenAI Gym by the way ? In addition, there are other continuous control task in OpenAI Gym, e.g. LunarLander, CarRacing, BipedalWalker, do you think we can try to use VIN on these tasks ? Perhaps even for some games in OpenAI Universe ? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#4>, or mute the thread https://github.com/notifications/unsubscribe-auth/AOeQNeIJyDZRiX2CxjebG1ux0sbTL4lQks5rLFisgaJpZM4LVG_I .
Could you release the Mars Navigation dataset used in your paper?
This is awesome, thank you so much!
Did you make mars.mat? Be able to share whether the code is reproducible?
Hello, I also got into trouble when I tried to use the 'obstacle_mars' class. It is not clear how to use it to generate data for training VIN? How is this generated to provide the following information?