Striker and Thrower xml not loading properly

[benelot]

There are some problems with loading the thrower.xml and the striker.xml with pybullet:
(https://github.com/bulletphysics/bullet3/blob/master/examples/pybullet/gym/pybullet_data/mjcf/striker.xml, https://github.com/bulletphysics/bullet3/blob/master/examples/pybullet/gym/pybullet_data/mjcf/thrower.xml)
-The thrower has some kind of fingers defined in the xml, however, they do not show up in the simulation.
-The striker xml contains a goal element, but it does not show up as 'goal'. Interestingly, the 'coaster' element shows up. Furthermore, the goal element is defined as 'free' to make it freely movable on the ground, which makes it float around strangely on the ground by itself and it falls half-way through the ground plane.

Any more details I should provide?

[erwincoumans]

If you want to help out, please consider submitting a patch to fix this. The MJCF importer is here:
https://github.com/bulletphysics/bullet3/blob/master/examples/Importers/ImportMJCFDemo/BulletMJCFImporter.cpp

Having a new port of the Ant and Half Cheetah (to start with) that behaves more similar to the original Gym MuJoCo environments (in addition of the existing Roboschool ports) would be higher on my wish list though.

[benelot]

For now, it is not on my list either. I just wanted to point out that there seems to be some discrepancy between the mujoco mjcf importer and the pybullet mjcf importer. So maybe at some point it can be fixed. But next up is definitely the Gym MuJoCo envs. I trained some agents on the original envs and now I am about to build the envs in pybullet. If I may call you in at some point to help me decrypt some of the mujoco internal values they are using as observations (basically stuff below --main inputs and outputs of the computation: http://www.mujoco.org/book/source/mjdata.h) , then that would be great. Some of them are needed to make similar observations from pybullet to the agent.

…

On Thu, May 31, 2018, 21:18 erwincoumans ***@***.***> wrote: If you want to help out, please consider submitting a patch to fix this. The MJCF importer is here: https://github.com/bulletphysics/bullet3/blob/master/examples/Importers/ImportMJCFDemo/BulletMJCFImporter.cpp Having a new port of the Ant and Half Cheetah (to start with) that behaves more similar to the original Gym MuJoCo environments (in addition of the existing Roboschool ports) would be higher on my wish list though. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1721 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AC97q0tzQlUvESR4G6G3cV24lPJsHUVTks5t4EINgaJpZM4UVi10> .

[benelot]

Hello Erwin,

I am working again on the mujoco envs and I get stuck with some mujoco related variables that are used as variables. Can you help me guess what they might correspond to in bullet?
Currently I am looking for:
self.sim.data.qfrc_constraint

In the above mentioned mjdata.h, it says the following about it:
mjtNum* qfrc_constraint; // constraint force (nv x 1)

In the double pendulum, this returns 3 different values. For now, I can live with replacing them with 0, but I would love to understand what they mean and what they correspond to.

How can I retrieve something like this in pybullet?