Run Ant-v1, got WARNING: Nan, Inf or huge value in QACC at DOF 0. The simulation is unstable. Time = 0.0000.

[andrewliao11]

Hi all, I run a3c in Ant-v1 and got this warning:

WARNING: Nan, Inf or huge value in QACC at DOF 0. The simulation is unstable. Time = 0.0000.

Are there anyone know what's the problem with this, and maybe the reason why. (My code runs well for other envs

[lhao499]

I met the same issue on several environments, does anyone know a solution?

[andrewliao11]

my mujoco-py version is mujoco-py==0.5.7 I'm not sure if it appears in other version

[kkjh0723]

@andrewliao11 I also tried Ant-v1 with your pytorch a3c code. I changed line 81 in test.py

state, reward, done, _ = env.step(action.numpy())

and then, the warning was disappeared

[HenryZhou7]

Any updates on this issue? I am having the same trouble. I suspect that it is due to some structure that causes the acceleration becoming too large.

[kirk86]

Same issue here with different environments. If anyone has found as solution plz kindly share!

[cyrilzakka]

Seems to be an issue with MuJoCo - too much penetration or contact between two bodies will cause this error. Model needs to updated

[muratcancicek]

I had a similar problem in #340. I discovered that you should slow down stepping. I solved the issue by just delaying the next step with time.sleep(.002).

time.sleep(.002) obs, r, done, _ = self.env.step(action)

[quantumiracle]

Similar problem when using this for a robotic control task.

[HankerSia]

Similar problem when using this for grasp task with a customized mesh file. Any new solution? I had tried the sleep solution, it seems no help...

[HYDesmondLiu]

Sleep solution is not helping for me, either.

[lucasjinreal]

any updates?

[MoritzLange]

I recommend taking a look at the action you're passing to step(). I'm debugging some quite complex code and at some point this error occured when the action just consisted of nans. So at least in my case the problem is exactly what the error message states.