Real-Time Physics Simulation Forum

Hi developers,

I trained a curiosity-driven PPO agent in Pybullet humanoid environment. Surprisingly, the humanoid finally learns "Flying" . It seems to some problems in Pybullet humanoid environment lead to such unrealistic behavior. The demo can be seen on my youtube page: https://www.youtube.com/watch?v=KrU34hSWH_E.

This is a major problem that has been obstructing my work as well. My robot goes flying into the air, violating the laws of physics, if it starts to fall over. I would really like to know if there's a way to mitigate this bug, so that I can code the robot properly.

I experience the same problem. I'm trying to make a quadruped jumping via reinforcement learning and he has learned flying. It's like it has learnt a way to use the actuators to generate forces to make it fly winning the gravity force.

Has someone found a workaround to this matter ?
Is there a way to change solver ?

We trained humanoid, it can have issues if you are not careful.

This can have various causes: hitting joint limits or using too large forces/torques, too large time step or other.

Try reducing the torques. Use a smaller time step.