Deadlock in "World" class

[Luocheng-Zheng]

Hello. Recently I am using raisimgym to train an RL policy. However, after several iterations, it got stuck. As the program uses parallel programming, the issue is like a deadlock: it stopped running forever, cannot be terminated by ctrl+c and CPU usage is low. I find out it was stuck in the line world->integrate();.

        mano_r_->setPdTarget(pTarget_clipped_r, vTarget_r_);
        /// Apply N control steps
        for (int i = 0; i < int(control_dt_ / simulation_dt_ + 1e-10); i++){
            if(server_) server_->lockVisualizationServerMutex();
            world_->integrate();
            if(server_) server_->unlockVisualizationServerMutex();
        }

Therefore, I wonder if there are any mutex or any other kind of locks in "World" class. Thank you in advance!

[jhwangbo]

Please do not post an image of code. Other people cannot search it.

Are you sure it's not stuck and the next line? integrate has no locking mechanism. Can you comment the line above and below and check if it gets stuck?

[Luocheng-Zheng]

Already tried. I am pretty sure that it was stuck in world_->integrate();

[jhwangbo]

can you create a minimum code to reproduce the issue? Because there is no locking function anywhere in the code, I don't know how I can help

[Luocheng-Zheng]

Thank you. Now at least I know that World does not produce that issue itself. I'll try to reproduce it.

[Luocheng-Zheng]

It turns out that my urdf model has inconsistent mass and inertia matrix, which may cause some problems in integrate.

[fanshi14]

cannot be terminated by ctrl+c and CPU usage is low

I have the very similar case for a long time, but do not find the reason. What do you mean that your urdf model has inconsistent mass and inertia matrix, could you share your urdf or minimal example of this kind of urdf? @Luocheng-Zheng

[Luocheng-Zheng]

@fanshi14 It is an articulated object, which has two bodies: lid and bottle. The wrong urdf accidentally set the bottle inertia the same as the lid, but kept the bottle mass. And it also had wrong rpy for <origin>. When you come accross similar problems about world->integrate(), do double check your urdf files.

[fanshi14]

Thanks for the information. @Luocheng-Zheng

I had the virtual links without any mass and inertia in my urdf; then I add a tiny ignorable mass/inertia value, the deadlock problem is less likely to happen (not 100% sure solved). Do you think it is a potential problem in raisim? @jhwangbo

[jhwangbo]

The zero-mass link is ok as long as it does not result in infinite acceleration. In this case, it will segfault or so. But I am not sure when it can get stuck. If you can provide a runnable code and the URDF, I can try to figure out the issue.

[ruanruan0313]

@fanshi14 It is an articulated object, which has two bodies: lid and bottle. The wrong urdf accidentally set the bottle inertia the same as the lid, but kept the bottle mass. And it also had wrong rpy for <origin>. When you come accross similar problems about world->integrate(), do double check your urdf files.

Hello, I would like to ask what to do if the replacement urdf file does not display the full model. thank you.