pybullet-gym
pybullet-gym copied to clipboard
benelot
Contributions
I am interested in contributing to this project. Do you want to lay out a more detailed roadmap for others who might be interested in making contributions?
Hey there! I am very happy that you are interested in contributing! The general theme of pybullet-gym is to free gym from the commercial mujoco engine. But it means that every gym must behave nearly equally to the original env in order to transfer the trained weights over from a mujoco env to the pybullet env. I will soon provide trained agents on the mujoco envs that can be used as benchmarks.
Here is a rough description of things that need to be done: -[ROBOSCHOOL GYMS] The current gyms are the roboschool gyms ported to pybullet. So far, most of them work well, except for the manipulator envs striker, pusher and thrower, where the robot is not correctly loaded. This will have to be fixed with Erwin Coumans, which is something I can report to him. -[OPENAI MUJOCO GYMS] Soon I will start to port the OpenAI gyms, which unfortunately have a slightly different observation (and probably action) vector. I can setup all the gyms quickly, but it will take a while to find out what some of observations are in mujoco and what they correspond to in pybullet. Some of the observations might not be exposed on pybullet, then we can request them, for others it is already hard to know what they are in mujoco. -[OPENAI ROBOTICS GYMS] Next in line would be the robotics gyms in OpenAI. These are particularly delicate simulations and might take some tuning to even be simulatable in pybullet. -[DEEPMIND CONTROL SUITE] Then there is Deepmind Control Suite, another set of gyms which are in mujoco and need to be freed. I did not look into those yet, but I am sure there are some tough parts too. -[STANFORD SUITE] https://github.com/StanfordVL/robosuite
I will turn this list into a cleaner roadmap on the readme page for others who might be interested in making contributions.
Do you have some experience with Mujoco? You can definitely be of help since there is a lot to be done! Thanks!
On Mon, May 28, 2018, 22:28 Jon Michaux [email protected] wrote:
I am interested in contributing to this project. Do you want to lay out a more detailed roadmap for others who might be interested in making contributions?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/benelot/pybullet-gym/issues/7, or mute the thread https://github.com/notifications/unsubscribe-auth/AC97q25Ptd29sjqfF_wqx5y5662RcZpJks5t3F3ogaJpZM4UQk7W .
Can you provide some better channel of communication, like on slack to get in touch more easily?
I am about to go on holidays. I will provide you with a slack channel by the beginning of october.
On Sat, Sep 15, 2018, 13:32 Novin Shahroudi [email protected] wrote:
Can you provide some better channel of communication, like on slack to get in touch more easily?
— You are receiving this because you commented.
Reply to this email directly, view it on GitHub https://github.com/benelot/pybullet-gym/issues/7#issuecomment-421554717, or mute the thread https://github.com/notifications/unsubscribe-auth/AC97q16Im7vp_4ER4l2lG92dDwoLuYwPks5ubOUxgaJpZM4UQk7W .
Hey there! I opened a slack channel for us to collaborate:
https://join.slack.com/t/pybullet-gym/shared_invite/enQtNTM2MDI3MTk0MzEwLTg1OTU5Y2ViNjQ4NWE1ZTQxMWEwMWNhYjMwYTc2ZTRkMzg5MWUyYmYwOGE3Njc3ODk1YzU2M2YyOTI0ODA3NmY
The most important thing to do at first is to rebuild all mujoco environments in pybullet. This is most easily done by taking the structure from the roboschool code and replacing inputs and outputs to match the inputs/outputs of an agent trained in mujoco.
If somebody has openai gym with mujoco installed, it would be great if somebody could train some tensorforce agents on it to perform well. Then we can connect them to the above built envs to tune them to behave similarly.
One of the main problems is that for some environment readings to the agent in mujoco I can not find the corresponding reading in pybullet. But maybe we can figure this out with the pybullet/bullet maintainer Erwin Coumans.
I can invest some time this week to do the first task.
Great! I need to look into installing MuJoCo. Currently the CPU on my research machine doesn't have the AVX instruction set required to run MuJoCo. I'll post any further updates to the slack channel.
what is the relationship between this repo and https://github.com/bulletphysics/bullet3/tree/master/examples/pybullet/gym ?
I am basically the maintainer of the envs in bullet3. However, since I have no push rights to bullet3 I prepare stuff here and then prepare a pull request. There was some discussion to make the implementations separate from the bullet3 repo, but for now they are kept in there.
On Sun, Nov 25, 2018, 04:09 Dan Brickley [email protected] wrote:
what is the relationship between this repo and https://github.com/bulletphysics/bullet3/tree/master/examples/pybullet/gym ?
— You are receiving this because you commented.
Reply to this email directly, view it on GitHub https://github.com/benelot/pybullet-gym/issues/7#issuecomment-441411938, or mute the thread https://github.com/notifications/unsubscribe-auth/AC97q_ofLmf_q0W_xPJTK9G666pyPkyEks5uygn5gaJpZM4UQk7W .
Since people are posting here, I want fetch support for my research, so I am in the middle of porting the fetch robot and will be doing extensive env setup. If anyone wants to help, I will be making branches on my fork, then doing pull requests to this repo.
Also the slack link is dead
Slack channel invite link is updated! Thanks for telling me, I forgot to turn off link expiration.
@benelot Any chance you could activate the slack invite again? Looking forward to using this env in the next period and I hope to contribute to it.
Slack channel invite link is updated! Thanks for telling me again. I think I turned off link expiration, but it seems to have expired nonetheless.