Train.py - model starts spinning even tho ang vel set to 0s, and then it shoots off horizontally?

[stuartc842]

i’m trying to design and train a model from scratch of a balancing bipedal robot

i made a simplified urdf model, and i tried simplifying ant.py from IsaacGymEnvs as much as possible as a template

i renamed my version from ant to antsimple and uploaded all the relevant files here GitHub - stuartcrobinson/isaac_gym_question: files to get help on why my simulation is so messed up

i would include some code here but i have no idea what’s responsible. why it’s not starting out motionless like the ant model does when i switch out my urdf for the original nx_ant.xml

very confused and any help or pointers would be greatly appreciated!

why does gravity seem to change axis half way through??

why does it start out spinning even tho i set velocities to 0?

self.initial_root_states[:, 7:13] = 0

python train.py task=Antsimple  test=True num_envs=1

[saner]

I sometimes obeserve similar behavior if convex decomposition results in a degenerate collision mesh. Some models will result in a spikey ball type collision mesh (which is usually pretty small so you’ll have too zoom quite a bit to see it) and so it just bounces around like crazy. In your case the motion seems less erratic than what I get, so observing the collision mesh to make sure it’s not degenerate will help eliminate the possibility.